Six Sigma is a data-based management and improvement methodology. Its distinctive contribution is not simply the use of statistical tools, but the disciplined conversion of operational observations into evidence for decisions. Descriptive statistics reveal what has happened; probability models describe what may happen; sampling theory supports conclusions about a larger population; confidence intervals express uncertainty; and graphical analysis exposes patterns that numerical summaries may conceal.

A Six Sigma practitioner must also distinguish ordinary, inherent process fluctuation from unusual, assignable disturbance. This distinction—between common-cause and special-cause variation—determines whether the correct response is process redesign or investigation of a specific event. Statistical knowledge therefore supports both the technical and managerial dimensions of DMAIC: Define, Measure, Analyze, Improve and Control.

Learning Objectives

1. Statistical Thinking in Six Sigma

Statistical thinking rests on three propositions. First, all work occurs in processes. Second, all processes exhibit variation. Third, understanding variation is necessary for improving performance. A metric such as average delivery time is incomplete unless its spread, distribution, measurement reliability and time order are also considered.

Concept	Meaning	Example
Population	Entire set of units or events of interest	All invoices processed this year
Sample	Subset observed from the population	150 selected invoices
Parameter	Numerical property of a population	Population mean μ
Statistic	Numerical property calculated from a sample	Sample mean x̄
Variable	Characteristic that can take different values	Processing time
Observation	One recorded value	42 minutes

A measurement may be continuous, such as weight or cycle time, or discrete, such as the number of defects. The scale of measurement and the process context influence the appropriate analysis.

2. Descriptive Statistics

2.1 Measures of location

Here, x̄ is the sample mean, xᵢ is observation i, and n is the sample size. The mean uses every observation and is appropriate for many approximately symmetric distributions, but it is sensitive to outliers.

The median is the middle ordered observation, or the average of the two middle observations when n is even. It is resistant to extreme values and is often preferable for skewed data such as waiting times. The mode is the most frequent value and is especially useful for categorical or discrete data.

2.2 Measures of dispersion

The range is the maximum minus the minimum. It is simple but depends only on two observations. The sample variance is:

The sample standard deviation is s = √s². The denominator n − 1, rather than n, provides an unbiased estimator of population variance under random sampling from a population with finite variance. Standard deviation has the same units as the original variable; variance has squared units.

Usually expressed as a percentage, it is useful for comparing relative variation across measures with different scales, provided the mean is positive and not close to zero.

Worked Example 1: Summary statistics

Five invoice-processing times are 8, 10, 10, 12 and 20 minutes. The mean is 60 ÷ 5 = 12 minutes. The median is 10 minutes. The range is 20 − 8 = 12 minutes. Deviations from the mean are −4, −2, −2, 0 and 8; their squared sum is 100. Thus:

The mean exceeds the median because the 20-minute observation creates right skew. Reporting only the mean would therefore provide an incomplete description.

2.3 Shape and robust summaries

Skewness describes asymmetry. Positive skew has a long right tail; negative skew has a long left tail. Kurtosis concerns tail weight and concentration relative to a reference distribution, although its interpretation should not be separated from graphical evidence.

For operational data, robust summaries such as the median and interquartile range (IQR) are valuable. The IQR is Q₃ − Q₁, where Q₁ and Q₃ are the first and third quartiles. A common exploratory rule flags values below Q₁ − 1.5(IQR) or above Q₃ + 1.5(IQR) as potential outliers. This is a screening rule, not proof that a value is erroneous or a special cause.

3. Probability and Distributions

Probability quantifies uncertainty. For an event A, P(A) lies between 0 and 1. The complement rule is P(not A) = 1 − P(A). For mutually exclusive events A and B, P(A or B) = P(A) + P(B). For general events, the intersection must be subtracted to avoid double counting.

A random variable assigns numerical values to outcomes. A discrete random variable has countable outcomes; a continuous random variable can take values over an interval. A probability distribution specifies the likelihood structure of a random variable through a probability mass function or probability density function.

3.1 Binomial distribution

The binomial model applies when there are a fixed number n of trials, each trial has two outcomes, trials are independent, and the probability p of success is constant. If X is the number of successes:

where k = 0, 1, …, n and C(n,k) = n! ÷ [k!(n − k)!]. The mean is np and the variance is np(1 − p). In quality work, “success” may mean a defect, a conforming unit or an accepted claim; the definition must be explicit.

3.2 Poisson distribution

The Poisson model describes counts of events in a fixed interval of time, area, length or volume when events occur independently at a stable average rate λ. For count X:

The mean and variance are both λ. The model is inappropriate when the event rate changes materially over the interval or events are strongly dependent. Overdispersion—observed variance greater than the mean—may indicate heterogeneity or dependence.

3.3 Normal distribution

A normal random variable has a symmetric, bell-shaped distribution characterized by mean μ and standard deviation σ. Standardization converts X to a z-score:

The z-score expresses distance from the mean in standard-deviation units. The empirical rule states that approximately 68%, 95% and 99.7% of observations lie within 1, 2 and 3 standard deviations of the mean, respectively, when the distribution is normal. These percentages are approximations, not universal laws.

The central limit theorem explains why sample means often become approximately normal as sample size increases, under suitable independence and finite-variance conditions, even when individual observations are not normal. It does not imply that every dataset is normal or that a small sample is automatically adequate.

3.4 Exponential and other useful models

The exponential distribution is often used for waiting times between Poisson events. Its rate parameter must be expressed in reciprocal time units. In reliability and service systems, Weibull, lognormal and gamma distributions may fit data better than the normal distribution. Distribution selection should be based on process knowledge, graphical diagnostics and formal assessment—not convenience alone.

4. Sampling Theory and Data Quality

A sample is useful only when its selection and measurement process support the intended inference. Simple random sampling gives each population unit a known, usually equal, selection probability. Stratified sampling divides the population into meaningful groups and samples within each; it can improve representation when groups differ. Systematic sampling selects every kth unit after a random start. Cluster sampling selects groups, often for logistical reasons, but may increase sampling variance because units within clusters resemble one another.

Convenience sampling is easy but vulnerable to selection bias. A sample collected only during day shifts, from willing customers or from successfully completed transactions may not represent the target population.

4.1 Sampling error and bias

Sampling error is random difference between a sample result and the population parameter. Bias is systematic deviation caused by the sampling or measurement process. Increasing sample size generally reduces random sampling error, but it does not remove systematic bias. A very large biased sample remains biased.

Time order is essential in process improvement. A random sample may estimate an overall population parameter, whereas a rational subgroup—observations collected close together under similar conditions—helps reveal within-process variation. Both perspectives may be required.

4.2 Measurement System Analysis

Observed variation combines process variation and measurement variation. If Y is the observed value, a simple conceptual model is:

where T is the true characteristic and E is measurement error. Measurement error may include repeatability (variation under the same operator and equipment) and reproducibility (variation across operators or conditions). Bias, linearity and stability are additional concerns. Before acting on small process differences, a Six Sigma team should establish that the measurement system is adequate for the decision.

5. Confidence Intervals and Statistical Inference

A point estimate gives one value; an interval estimate communicates uncertainty. For a population mean with known σ, a confidence interval is:

Here α is the significance level, z₍α⁄₂₎ is the standard-normal critical value leaving α⁄2 in each tail, σ is population standard deviation and n is sample size. The formula assumes independent observations and a normal population or sufficiently large sample for the sampling mean to be approximately normal.

where t is the Student’s t critical value with n − 1 degrees of freedom and s is the sample standard deviation. For a population proportion p, an approximate large-sample interval is:

This approximation requires adequate expected counts; for small samples or proportions near 0 or 1, exact or improved intervals such as Wilson intervals are preferable.

A 95% confidence interval does not mean there is a 95% probability that a fixed parameter lies in this particular interval. Under repeated sampling and valid assumptions, approximately 95% of intervals constructed by the method would contain the parameter.

Worked Example 2: Mean confidence interval

A sample of 25 cycle times has x̄ = 42 seconds and s = 5 seconds. For a 95% interval with 24 degrees of freedom, t ≈ 2.064. The margin of error is:

Therefore, the interval is 42 ± 2.064, or approximately 39.94 to 44.06 seconds. This interval estimates the long-run mean only if the sample is reasonably representative and the independence and distributional assumptions are defensible.

Confidence intervals support practical decisions, but statistical significance is not the same as managerial importance. A tiny improvement can be statistically detectable in a huge sample yet irrelevant to customers or cost. Conversely, a practically important change may be uncertain in a small sample.

6. Graphical Analysis

Graphs are analytical instruments, not decoration. A histogram shows the distribution of a continuous measure; it can reveal skewness, multiple peaks, gaps and unusual tails. A box plot displays median, quartiles, IQR and potential outliers, and supports group comparison. A scatter plot examines the relationship between two quantitative variables; correlation alone does not establish causation.

A run chart plots observations in time order and helps detect shifts, trends, cycles and mixtures. A control chart adds statistically derived limits and is discussed below. A Pareto chart ranks categories by frequency or impact; its descending bars and cumulative line help prioritize, but the “vital few” principle is not a guarantee that a fixed percentage of causes creates a fixed percentage of effects.

Stratification is often more revealing than aggregation. A combined histogram may appear stable while separate distributions by supplier, machine, product family or shift show major differences. Every graph should identify the measure, units, time period, subgroup and sample size.

7. Common-Cause and Special-Cause Variation

Common-cause variation is produced by the routine design and operation of a process: methods, materials, equipment, environment and management systems. It forms the process’s current pattern of variation. Special-cause variation arises from an identifiable, unusual influence, such as a broken sensor, incorrect setup, contaminated material or exceptional staffing event.

The distinction is contextual. A cause may be special relative to one process state but common after it becomes embedded in the routine system. A point outside a control limit is evidence to investigate, not automatic proof of a particular cause.

7.1 Control charts

A basic individuals chart uses a process center line and limits commonly estimated as:

For a chart of individual observations, σ should be estimated appropriately, often from moving ranges rather than treating observations as independent subgroup means. For an X̄ chart based on rational subgroups, control-limit constants depend on subgroup size and the selected chart design. Control limits are not specification limits: control limits describe process behavior, whereas specification limits express customer, engineering or regulatory requirements.

Signals such as a point beyond a limit, a sustained shift, a trend or an unusual nonrandom pattern suggest special-cause investigation. Rules must be defined before analysis where possible, because adding many rules after viewing data increases false alarms. Autocorrelation, seasonality and changing measurement systems can also create misleading signals.

7.2 Correct response

When special cause is suspected, contain potentially affected output, verify the measurement, examine the time and conditions of the signal, identify the assignable cause, and document corrective action. When only common-cause variation is present, repeated adjustment of individual outputs—tampering—usually increases variation. Improvement requires changing the system: redesigning the process, reducing input variation, improving maintenance, revising standard work or changing the operating conditions.

8. Illustrative Case: Reducing Laboratory Turnaround Time

Background and problem

The following is a fictional illustrative case. A hospital laboratory promises routine chemistry results within 60 minutes. A DMAIC team observes 40 consecutive turnaround times in minutes and finds a mean of 58, a median of 54 and a standard deviation of 18. The right-skewed pattern includes several results above 100 minutes. Management initially proposes disciplining the staff member associated with the slowest result.

Analysis

The team first validates timestamps and discovers that collection-to-receipt and receipt-to-verification had been inconsistently defined. After operational definitions are standardized, a time-ordered plot is stratified by shift and analyzer. The data show a stable pattern on two analyzers but a late-shift cluster of long times on a third. A box plot by analyzer confirms higher variation for that analyzer. A measurement-system check finds that the laboratory information system records verification time reliably, but manual receipt entry is delayed during specimen batching.

The team avoids treating the overall mean as sufficient evidence. It separates routine process behavior from a special event: a reagent-temperature alarm explains one extreme observation, while batching and staffing allocation explain the recurring late-shift pattern. The alarm is special cause; the staffing and workflow design are common causes of the recurring process state.

Approach and outcome

The laboratory repairs the temperature-control unit, introduces a standardized electronic receipt scan, changes late-shift staffing and establishes a control chart for verified turnaround time. After implementation, the mean falls to 49 minutes and the standard deviation to 9 minutes in a comparable monitoring period. The team reports a confidence interval for the mean, checks that the measurement definition has remained unchanged, and monitors the chart rather than declaring success from a single favorable week.

Learning points

9. Limitations, Governance and Contemporary Practice

Statistical tools can be misused through p-hacking, selective exclusion of inconvenient observations, repeated subgroup testing, or treating correlation as causation. Data privacy matters when process records contain patient, employee or customer information. Access should be limited to legitimate purposes, personally identifiable information should be minimized, and analytical decisions should be auditable.

Automated analytics and AI can assist with anomaly detection, image-based inspection, forecasting and high-dimensional pattern recognition. However, models may reproduce biased data, confuse a data-quality problem with a process problem, or generate false alarms when the process changes. Human review, version control, explainable decision rules and ongoing validation remain necessary. AI should augment—not replace—process knowledge, measurement-system analysis and responsible managerial judgment.

Best practice is to define the question and unit of analysis first; verify data provenance and measurement capability; visualize before modeling; preserve time order; state assumptions; distinguish statistical from practical significance; triangulate numerical and contextual evidence; and document decisions sufficiently for replication.

Chapter Summary

Statistical foundations enable Six Sigma teams to describe processes, quantify uncertainty and act appropriately on variation. Mean, median, variance, standard deviation and robust measures summarize location and spread, but must be interpreted with distributional shape and time order. Binomial, Poisson, normal and other distributions provide useful models only when their assumptions match the process. Sampling design determines whether conclusions can be generalized, while measurement-system analysis determines whether observed differences are credible.

Confidence intervals express estimation uncertainty and should not be confused with probabilities about a fixed parameter or with practical importance. Graphical tools reveal skewness, stratification, trends and unusual patterns. Most importantly, common-cause variation calls for systemic improvement, whereas special-cause variation calls for investigation and correction. Sound Six Sigma practice combines statistical discipline with process knowledge, ethical governance and sustained control.

Key Terms

Variance: Average squared dispersion measure, using an appropriate population or sample denominator.

Coefficient of variation: Relative dispersion, calculated as standard deviation divided by mean.

Probability distribution: Model specifying likelihoods for values of a random variable.

Binomial distribution: Distribution for successes across fixed independent two-outcome trials.

Confidence interval: Interval produced by a method designed to contain a parameter at a stated long-run frequency.

Sampling error: Random difference between a sample estimate and population parameter.

Measurement-system analysis: Evaluation of measurement error and measurement capability.

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

Special-cause variation: Unusual variation attributable to an identifiable influence.

Control limits: Statistically derived boundaries describing expected process behavior.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

Question 2

For a sample of 25 observations with x̄ = 42 and s = 5, using t ≈ 2.064, what is the approximate 95% confidence interval for the mean?

Question 3

Question 4

Question 5

Question 6

Which sampling method divides a population into meaningful groups and samples within each group?

Question 7

A sustained process shift caused by a new, recurring staffing arrangement is best classified as:

Chapter 6 Statistical Foundations for Six Sigma

Statistical Foundations for Six Sigma

Introduction