Lead DMAIC decisions with reliable measurement, defensible statistical inference and controls that sustain improvement. This guide combines twenty original multiple-choice questions with sixty worked concepts covering project selection, process characterization, measurement systems, capability, hypothesis testing, regression, designed experiments and SPC. It follows the supplied PeopleCert IASSC ICBB English syllabus, Version 1.1, January 2026, for study as of 16 October 2026; no future syllabus is incorporated. The preview allocates questions 3/4/6/3/4 and concepts 8/13/16/10/13 across Define, Measure, Analyze, Improve and Control. These are learning allocations, not official examination counts. Citations establish syllabus relevance; worked calculations and technical explanations are original instruction.
What this guide covers
Master ICBB DMAIC with 20 original practice questions and 60 lessons on statistics, DOE and control.
Define Phase
1. DMAIC and the causal model Y = f(X)
DMAIC develops an evidence-based explanation of how process inputs influence an output. Define establishes the business problem; Measure establishes trustworthy baseline data; Analyze evaluates suspected causes; Improve changes influential inputs; Control sustains the resulting process. In Y = f(X), Y is the response and X represents potentially controllable or contextual inputs. The equation expresses a causal investigation strategy, rather than proving that an observed association is causal. Six Sigma developed through industrial variation-reduction work, while modern applications include service processes. Black Belts coordinate the technical investigation across phases.
Worked example: A repair center defines Y as elapsed days from receipt to return. It measures queue time, parts availability and repair complexity as candidate X variables, then tests whether changing parts replenishment reduces delays without increasing repeat repairs.
Common mistake: Treating a correlated input as a proven cause skips the design, measurement and alternative-explanation checks needed for credible improvement.
Official syllabus reference: [1]
2. Voice-to-CTQ translation and operational definitions
Voice of the customer describes desired outcomes, voice of the business expresses organizational requirements, and voice of the employee reveals practical operating knowledge. A CTQ translates a relevant need into a measurable characteristic with defined units, boundaries and acceptance rules. Separate a broad need, its performance driver and its measurable requirement. Business cost and employee workload can become balancing measures when improving a customer outcome. Conflicting voices require explicit trade-offs rather than silently replacing customer value with an easy internal productivity measure. An operational definition ensures different observers classify the same event consistently.
Worked example: Customers want invoices that are easy to resolve. The team defines a CTQ as the percentage of disputed invoices resolved within five agreed business days, specifies when the clock starts, and tracks employee overtime as a balancing measure.
Common mistake: Using training attendance as the CTQ confuses an improvement activity with the customer-facing result the activity is intended to influence.
Official syllabus reference: [1]
3. Project charters, scope boundaries and ownership
A charter connects an observed performance gap to strategic value, a bounded process and accountable ownership. Its problem statement describes what happens, where, when and how much, without embedding an unverified cause or preferred solution. The goal specifies a measurable improvement and timeframe, supported by primary and balancing metrics. Scope identifies included and excluded process boundaries. The sponsor supplies authority and removes barriers; the Black Belt leads analysis; subject specialists contribute process knowledge; the process owner accepts ongoing operation. These roles prevent analytical responsibility from being confused with ownership of sustained performance.
Worked example: A charter targets reducing median warranty-claim processing time from twelve to eight days within six months. It includes intake through authorization, excludes product redesign, names the claims director as process owner, and balances speed against incorrect approvals.
Common mistake: Writing 'install automation to fix slow claims' as the problem statement commits to a solution before its causal justification is established.
Official syllabus reference: [1]
4. COPQ, incremental cash flow and discounted benefits
Cost of poor quality includes costs arising from failures, such as scrap, rework, complaints and warranty claims. Prevention and appraisal are separate cost-of-quality categories, although a project may change them. Financial evaluation uses incremental effects relative to an explicit baseline. Released staff capacity becomes cash savings only when spending actually falls, or becomes another benefit through a separately justified use. Net present value discounts each future net cash flow and subtracts investment. Benefits capture assigns ownership, timing and verification rules, preventing overlap between reduced rework, increased capacity and additional contribution.
Worked example: A project removes $80,000 of annual scrap spending and requires $15,000 of annual maintenance. Its recurring net cash benefit is $65,000. The team records released inspection hours separately until a credible additional use or spending reduction is demonstrated.
Common mistake: Counting the same labor hours as both payroll savings and capacity supporting additional sales inflates benefits through double counting.
Official syllabus reference: [1]
5. Defect opportunities, DPU and DPMO
A defect is a failed requirement; a defective unit has at least one defect. DPU equals total defects divided by units inspected and can exceed one because a unit can contain multiple defects. DPMO equals defects divided by total defined opportunities, multiplied by one million. With a constant opportunity count per unit, opportunities equal units times opportunities per unit. Comparisons require consistent, meaningful opportunity definitions and inspection coverage. Adding trivial opportunities artificially lowers DPMO without improving output. Converting DPMO into a sigma label requires an explicit distributional and shift convention rather than an automatic universal rule.
Worked example: Inspectors find 45 defects across 300 forms, each with five predefined requirements. DPU is 45/300 = 0.15 and DPMO is 45/1,500 × 1,000,000 = 30,000. These figures do not identify how many forms had at least one defect.
Common mistake: Reporting 45 defective forms from 45 defects assumes every defect occurred on a different form, which the count does not establish.
Official syllabus reference: [1]
6. First-time yield, RTY and hidden rework
First-time yield is the fraction of units passing a defined step without repair, rework or repeat processing. Rolled throughput yield multiplies conditional first-time yields across sequential steps, estimating end-to-end passage without failure. Final yield can remain high when extensive rework recovers failures, so it answers a different question. Define the unit, route and treatment of repeated visits consistently, especially when routes branch. Cycle time measures elapsed time for a defined operation or process boundary; distinguish hands-on processing from total elapsed lead time when waiting is substantial. Yield and time jointly reveal hidden operating burdens.
Worked example: Three approval steps have conditional first-time yields of 0.95, 0.90 and 0.98. Their RTY is 0.8379, even if nearly every application eventually finishes. Repeated corrections explain why final completion looks strong while customers still experience long delays.
Common mistake: Averaging step yields does not produce an end-to-end first-pass probability and can conceal substantial cumulative rework across a long process.
Official syllabus reference: [1]
7. Pareto analysis by frequency and economic impact
A Pareto chart orders categories from largest to smallest and often adds cumulative percentage. It focuses investigation on concentrated contributors, but the familiar 80:20 relationship is a heuristic rather than a required data pattern. Choose a measure aligned with the decision: occurrence counts identify frequent problems, while cost, delay or severity may produce different priorities. Categories need clear definitions and comparable observation periods. Exposure differences may require rate comparisons before attributing poor performance. Pareto ranking generates priorities for investigation; it does not establish root causes or eliminate the need to address severe low-frequency events.
Worked example: A service desk records 120 password incidents costing $600 and 30 integration incidents costing $6,000. Frequency ranking points toward passwords, while cost ranking points toward integration. The project sponsor chooses the cost view because reducing support expenditure is the stated objective.
Common mistake: Assuming the tallest frequency bar must deliver the greatest financial benefit ignores differences in cost per event and feasible improvement.
Official syllabus reference: [1]
8. Lean waste classification and the 5S sequence
Lean, strongly associated with Toyota's production system, improves flow by removing work that consumes resources without creating customer value. The detailed syllabus lists overproduction, correction, inventory, motion, overprocessing, conveyance and waiting. Unused talent is also discussed in the supplied introductory material as an additional waste. Distinguish motion of people from transportation of items, and excess processing from correction of failures. The 5S sequence sorts unnecessary items, arranges needed items, cleans and inspects, establishes standards, and sustains discipline. Lean flow improvements complement Six Sigma's reduction of variation and defects.
Worked example: An operator walks repeatedly to a shared tool cabinet while unfinished work accumulates beside the station. The team identifies motion and inventory waste, removes unused tools, assigns visible locations for needed tools, and checks that the new arrangement reduces searching.
Common mistake: Calling every inspection waste to remove immediately overlooks whether process reliability currently requires that appraisal activity to protect the customer.
Official syllabus reference: [1]
Measure Phase
9. SIPOC boundaries versus detailed process maps
SIPOC summarizes suppliers, inputs, major process steps, outputs and customers. Its purpose is to agree on the process boundary and key interfaces before collecting detailed data. A detailed process map then represents decisions, loops, handoffs and actual work sequences within that boundary. Walk the process rather than relying only on the documented procedure, because informal rework loops often explain delay and variation. Mapping symbols organize understanding but do not prove performance. Operational definitions for starts, finishes and exceptions connect the map to measurable events and prevent inconsistent baseline denominators.
Worked example: For equipment returns, a SIPOC starts with the customer's return request and ends with refund confirmation. A detailed map then reveals an undocumented approval loop between service and finance, allowing the team to timestamp that loop separately.
Common mistake: Expanding a SIPOC into dozens of task boxes obscures its boundary-setting purpose and makes agreement on major interfaces unnecessarily difficult.
Official syllabus reference: [1]
10. Value-stream timing, takt and Little's law
A value-stream map joins material or work flow with information flow, queues and timing. Compare processing time with total lead time to expose waiting and inventory. Takt time equals available production time divided by customer demand for the same period; it expresses the required demand pace, rather than observed processing speed. For a stable flow system using consistent boundaries and long-run averages, Little's law links work in process, throughput and lead time: WIP = throughput × lead time. These relationships support diagnosis, but unstable backlogs or mismatched units make naive calculations misleading.
Worked example: A cell has 420 available minutes and daily demand of 210 units, giving a two-minute takt. With average WIP of 600 units and throughput of 200 units daily, Little's law implies three days of average lead time under stable-flow assumptions.
Common mistake: Treating takt time as a measured cycle time confuses the customer-demand requirement with the process's demonstrated ability to produce output.
Official syllabus reference: [1]
11. Fishbone hypotheses and X–Y prioritization
A fishbone diagram organizes plausible causes of an observed effect, often under categories such as methods, machines, materials, people, measurement and environment. Its branches are hypotheses, not verified causes. An X–Y matrix prioritizes candidate inputs by rating their relationships to important outputs and weighting those outputs by customer or business importance. The resulting scores guide measurement effort when investigating every input is impractical. Ratings remain judgment-based and can be influenced by dominant participants. Preserve a route for testing low-ranked possibilities when data contradict the initial theory.
Worked example: A team investigating late shipments lists scheduling, missing materials and scanner errors on a fishbone. An X–Y matrix ranks material availability highly because it affects both completeness and delivery timing, so the team first collects shortage and delay timestamps.
Common mistake: Declaring the highest-scoring X–Y input the root cause mistakes a structured prioritization judgment for evidence that changing it changes the output.
Official syllabus reference: [1]
12. FMEA effects, causes and detection controls
Process FMEA distinguishes a failure mode, its downstream effect, its potential cause and the controls that prevent or detect it. Severity describes the consequence, occurrence concerns the cause or failure likelihood under the chosen scale, and detection concerns the ability of current controls to find the problem before escape. Traditional RPN multiplies these ratings, but equal products can hide very different risk profiles. Prioritize severe effects explicitly and use the organization's documented method rather than inventing universal cutoffs. Actions should target prevention or detection mechanisms, followed by evidence-based reassessment of the affected ratings.
Worked example: A filling operation can underfill because a valve sticks. The effect is insufficient customer quantity, while a downstream weight check is a detection control. Replacing the valve mechanism addresses occurrence; improving the weight check addresses detection rather than inherent severity.
Common mistake: Reducing severity after adding an inspection confuses improved detection with a reduction in the consequence if the failure reaches the customer.
Official syllabus reference: [1]
13. Location, spread and robust descriptive statistics
The mean uses every observation and is sensitive to extreme values; the median identifies the middle ordered value and is more resistant to extremes. Sample variance averages squared deviations using n − 1, and standard deviation restores the original units. The interquartile range summarizes the middle half of ordered observations. Choose summaries that reflect the distribution and operational question, rather than reporting a mean alone. Weighted means require valid exposure weights. Descriptive statistics summarize the observed sample; population conclusions additionally require attention to sampling, dependence and estimation uncertainty.
Worked example: Five repair times are 2, 3, 3, 4 and 28 hours. Their mean is eight hours, while their median is three hours. Reporting both reveals the long repair that dominates the average and prompts investigation of its process circumstances.
Common mistake: Replacing every extreme observation with the median can erase legitimate process behavior and understate the actual variability customers experience.
Official syllabus reference: [1]
14. Normality assessment with plots and time order
A normal distribution is symmetric and characterized by its mean and standard deviation, but many process outputs are skewed, bounded or mixtures. Histograms reveal shape, probability plots reveal systematic departures, and time plots reveal shifts that pooled distributions can hide. A normality-test p-value depends on sample size: small samples may miss important departures, while large samples detect minor ones. Assess normality for the quantity required by the method, such as residuals or paired differences. Approximate normality does not establish independence or stability, and a transformation cannot automatically repair those separate problems.
Worked example: A pooled histogram of shaft measurements looks roughly bell-shaped, but a time plot reveals three shifts following tool changes. The team separates the operating periods before deciding whether a single normal capability model meaningfully represents the process.
Common mistake: A large normality-test p-value does not prove normality, particularly when the sample is small or observations were pooled across shifting conditions.
Official syllabus reference: [1]
15. Precision and accuracy answer different questions
Precision concerns the spread of repeated measurements under specified conditions. Accuracy concerns agreement with an accepted reference and is affected by systematic error as well as random variation. A tightly clustered measurement system can be consistently wrong, while an unbiased average can hide excessive reading-to-reading scatter. Define the operating conditions, reference quality and relevant measurement range before evaluating either property. Instrument display resolution is another distinct issue: many displayed decimal places do not establish either low variation or low bias. Measurement quality must be adequate for the process decision being made.
Worked example: A balance gives repeated readings between 100.18 and 100.22 grams for a certified 100.00-gram mass. Its repeatability appears strong at that condition, but the approximately 0.20-gram offset can distort decisions near an acceptance boundary.
Common mistake: Assuming extra decimal places establish accuracy confuses display resolution with demonstrated agreement to a trustworthy reference under actual operating conditions.
Official syllabus reference: [1]
16. Bias, linearity and stability in measurement
Bias is the difference between the expected measurement and an accepted reference at a specified level. Linearity concerns how that bias changes across the operating range. Stability concerns whether measurement performance changes over time under comparable conditions. Study bias with repeated reference measurements, linearity with references spanning the range, and stability with repeated checks in time order. Reference uncertainty and environmental differences affect interpretation. Correcting a constant offset may address local bias, but it does not establish linearity or remove drift. Each property requires evidence matched to its own question.
Worked example: A caliper reads 0.01 mm high on a small reference and 0.09 mm high on a large reference. A single offset correction would leave size-dependent error, so the team evaluates calibration across the working range before accepting routine measurements.
Common mistake: A successful calibration check at one size and one time does not establish reliable performance across the full range or future periods.
Official syllabus reference: [1]
17. Crossed gage R&R and variance decomposition
A crossed gage R&R study has multiple appraisers repeatedly measure the same parts, typically with randomized order and concealed prior results. Repeatability captures variation under the same appraiser and equipment conditions; reproducibility captures appraiser-related differences, potentially including part-by-appraiser interaction. Independent variance components add, whereas standard deviations do not. Percentage contribution compares variances; percentage study variation compares standard deviations. Part selection should represent the intended process range. Destructive measurement may require a nested design instead. Acceptance depends on the measurement use and specified criteria, rather than a universal threshold inferred from the syllabus.
Worked example: Repeatability variance is 0.006 mm² and reproducibility variance is 0.003 mm². Combined gage standard deviation is √0.009, approximately 0.0949 mm. The analyst compares that result with process spread or tolerance using the explicitly chosen reporting convention.
Common mistake: Adding repeatability and reproducibility standard deviations overstates combined variation when the appropriate independent components are expressed as variances.
Official syllabus reference: [1]
18. Attribute agreement, reference truth and kappa
Attribute MSA evaluates consistency of classifications rather than numerical readings. Repeated blind classifications assess within-appraiser agreement; comparisons across appraisers assess reproducibility; comparisons with defensible reference classifications assess correctness. Overall agreement can hide an important imbalance between false acceptance and false rejection. Kappa adjusts observed agreement for an estimated chance-agreement level, but its interpretation is sensitive to category prevalence and marginal classification patterns. Include representative borderline cases and enough examples of important failure classes. High agreement among appraisers does not demonstrate accuracy if all apply the same mistaken rule.
Worked example: Two inspectors agree on 98 of 100 products, but a reference review shows both accepted six known defective items. Their mutual agreement is high, while their false-acceptance performance reveals a weakness in the classification procedure or training.
Common mistake: Reporting only overall agreement can conceal poor detection of a rare but important defect class within a sample dominated by acceptable products.
Official syllabus reference: [1]
19. Cp and Cpk separate spread from centering
For an appropriate stable normal process model, Cp compares specification width with six within-process standard deviations: Cp = (USL − LSL)/(6σwithin). Cpk also accounts for location, taking the smaller of (USL − mean)/(3σwithin) and (mean − LSL)/(3σwithin). Thus Cpk cannot exceed Cp when calculated consistently. A high Cp with lower Cpk suggests centering opportunity, while low Cp indicates excessive short-term spread relative to tolerance. Specifications come from requirements, not control-chart calculations. Capability conclusions also depend on measurement adequacy, representative data and justified distribution assumptions.
Worked example: A stable process has limits of 20 and 26 mm, mean 24 mm and within-process standard deviation 0.5 mm. Cp is 6/3 = 2.00, but Cpk is min(2,4)/1.5 = 1.33 because the upper limit is nearer.
Common mistake: Using the farther specification distance overstates capability because the nearer boundary determines the greater immediate risk of nonconforming output.
Official syllabus reference: [1]
20. Pp, Ppk and the stability boundary
Pp and Ppk use the overall sample standard deviation, whereas Cp and Cpk conventionally use an estimate of within-subgroup variation. Overall spread can include drift, shifts and between-subgroup differences, so performance indices may be lower than corresponding within-process indices. The distinction depends on the estimation method and sampling design, not merely on calling one dataset short-term and another long-term. An unstable process can still produce numerical indices, but a single distribution may poorly predict future output. Inspect time order and subgroup behavior before treating historical performance as demonstrated stable capability.
Worked example: A filling process has similar within-hour spread each day, but its hourly averages move after temperature changes. Cpk looks favorable while Ppk is lower. The team investigates the temperature-related shifts rather than claiming the within-hour result describes overall customer exposure.
Common mistake: A favorable within-subgroup index does not cancel the customer impact of between-subgroup shifts that occur during ordinary production.
Official syllabus reference: [1]
21. Discrete capability and exposure-adjusted defect rates
Attribute capability commonly describes the proportion nonconforming or a defect rate per defined exposure. Binary unit outcomes suggest a binomial model, while counts of defects may suggest a Poisson model when its assumptions are credible. A unit can contain several defects, so defect rate and defective proportion are not interchangeable. Varying inspected volume or opportunity requires explicit denominators. Report uncertainty, particularly when failures are rare; observing zero failures does not establish zero population risk. Check stability and possible clustering or overdispersion before interpreting a pooled rate as a predictable process characteristic.
Worked example: An audit finds twelve errors in 600 transactions, but only nine transactions contain any error. Defects per transaction are 0.020, while the observed nonconforming proportion is 0.015. Each answers a different question about customer exposure and correction workload.
Common mistake: Treating a defect count as a count of defective units overstates the nonconforming proportion whenever multiple defects occur within the same unit.
Official syllabus reference: [1]
Analyze Phase
22. Multi-vari analysis separates variation sources
Multi-vari analysis structures observations to compare variation within a unit or position, between units and across time. Its value comes from preserving the sampling hierarchy rather than pooling every measurement into one overall standard deviation. Large within-unit differences suggest different causes from large time-to-time shifts. The study is diagnostic and observational, so visible patterns generate focused hypotheses rather than proving causality. Sampling should span relevant positions, consecutive units and operating periods. Measurement-system variation must be understood, because an unreliable instrument can imitate or obscure the variation components being investigated.
Worked example: A team measures wall thickness at four positions on each of three consecutive tubes during several shifts. Differences within tubes dominate, while shift averages remain similar, directing investigation toward circumferential forming conditions rather than shift-to-shift setup practices.
Common mistake: Pooling all readings before plotting their hierarchy can hide whether the dominant variation occurs within products, between products or over time.
Official syllabus reference: [1]
23. Distribution families follow the data-generating process
Distribution choice begins with what is measured and how it arises. Binomial models describe a count of successes in a fixed number of independent trials with common probability. Poisson models describe event counts over exposure under a suitable rate process. Normal models describe symmetric continuous variation, while lognormal or other positive distributions can represent right-skewed durations or sizes. Exponential waiting times require a constant-rate memoryless process. These are modeling assumptions, not labels selected solely from a histogram. Mixtures, dependence and changing conditions can invalidate a familiar family even when the visual fit looks reasonable.
Worked example: A help desk counts arrivals during equal ten-minute periods and separately measures each customer's waiting time. A Poisson model might describe arrivals, while a positive skewed model may describe waits; the same distribution is not automatically appropriate for both variables.
Common mistake: Selecting a normal model merely because data are numerical overlooks count structure, physical bounds, skewness and the mechanism producing the observations.
Official syllabus reference: [1]
24. Sampling designs, stratification and selection bias
Simple random sampling gives population units known equal selection chances. Stratified sampling samples within meaningful groups and can improve coverage or precision; population estimates must reflect appropriate stratum weights. Cluster sampling selects groups of units and requires attention to within-cluster dependence. Systematic sampling can be efficient, but a periodic process pattern may align with the sampling interval. Convenience samples are vulnerable to selection bias. Increasing the size of a biased sample does not repair its representativeness. Define the target population, sampling frame, unit and time coverage before choosing the method.
Worked example: A company handles 80% routine and 20% complex claims but audits equal numbers from both groups. To estimate the overall defect proportion, it weights the two stratum estimates by 0.80 and 0.20 rather than averaging them equally.
Common mistake: Treating deliberately equal stratum samples as equal population shares biases the overall estimate when the underlying process volumes differ substantially.
Official syllabus reference: [1]
25. The central limit theorem and standard error
Under suitable independence and finite-variance conditions, the sampling distribution of a mean becomes approximately normal as sample size grows, even when individual observations are nonnormal. The standard error of an independent-sample mean is σ/√n, estimated using s when appropriate. This concerns repeated sample means, not a transformation of the raw observations into normal data. Required sample size depends on skewness, tails and dependence; there is no universal number that guarantees adequacy. Serial correlation or clustering changes effective information, and heavy-tailed behavior can make normal approximation unreliable despite a seemingly substantial sample.
Worked example: Individual service times are right-skewed with standard deviation twelve minutes. For independent samples of 36 services, the estimated standard error of the sample mean is twelve divided by six, or two minutes, while individual times remain skewed.
Common mistake: Claiming that a large sample makes the original observations normally distributed confuses a statistic's sampling distribution with the population's actual shape.
Official syllabus reference: [1]
26. Confidence intervals and the target of inference
A confidence interval combines an estimate with uncertainty from a specified sampling model. In repeated use, a valid 95% interval procedure covers the fixed population parameter in approximately 95% of comparable samples. It does not say that 95% of individual observations fall inside the calculated interval. Width depends on variability, sample size, confidence level and design. State the parameter explicitly, such as a mean difference or defect-proportion difference. Generalization also depends on sampling and study design; a narrow interval around a biased estimate can be precise without being accurate.
Worked example: A study estimates a mean delay reduction of four minutes with a 95% confidence interval from one to seven minutes. The interval concerns the population mean reduction; it does not promise that each individual transaction becomes one to seven minutes faster.
Common mistake: Interpreting a confidence interval as the range containing most individual outcomes ignores the additional variation between individual observations.
Official syllabus reference: [1]
27. Alpha, beta, power and prespecified alternatives
A Type I error rejects a true null hypothesis; alpha is the test's specified Type I error rate under its assumptions. A Type II error fails to reject the null at a particular alternative; beta depends on the actual effect, variability, sample size and test. Power equals 1 − beta. Choose one-sided or two-sided alternatives before examining results, based on the decision question. Lower alpha generally reduces power at fixed sample size, while larger samples improve sensitivity. A p-value measures incompatibility with the null model, rather than the probability that the null is true.
Worked example: A team plans 90% power to detect a two-percentage-point defect reduction at alpha 0.05. This targets a 10% miss probability at that specific reduction; it does not imply 90% power for every smaller improvement.
Common mistake: Reading p = 0.03 as a 3% probability that the null hypothesis is true assigns the p-value a meaning it does not have.
Official syllabus reference: [1]
28. Statistical significance versus practical importance
Statistical significance assesses evidence against a null model, while practical importance assesses whether an effect matters operationally. Large samples can make tiny differences statistically detectable; small studies can leave meaningful effects uncertain. Define a practically important difference before analysis and interpret effect estimates with confidence intervals against that criterion. A nonsignificant superiority test does not establish equivalence. Equivalence requires a suitable procedure and prespecified acceptable bounds. Likewise, a significant difference does not establish an attractive financial return or acceptable customer outcome. Statistical and business judgments need related but distinct evidence.
Worked example: A very large study estimates a cycle-time reduction of 0.2 seconds with a narrow interval excluding zero. If the business needs at least ten seconds to change capacity, the result is statistically clear but insufficient for the intended operational benefit.
Common mistake: Treating a nonsignificant result as proof of no meaningful difference overlooks effects that the study may lack enough precision to exclude.
Official syllabus reference: [1]
29. One-sample, independent and paired t-tests
A one-sample t-test compares a population mean with a specified reference using the sample mean, standard deviation and sample size. Independent two-sample procedures compare unrelated groups; Welch's version does not require equal population variances. A paired t-test analyzes within-pair differences and uses their standard deviation, preserving matching. Normality considerations apply to the relevant errors or differences, especially in small samples. Independence between experimental units remains essential. Choose the test from the study design and target parameter, rather than selecting whichever procedure produces the smallest p-value.
Worked example: The same twelve pumps are tested with two settings. The analyst calculates one difference per pump and tests the mean of those twelve differences against zero, using eleven degrees of freedom rather than treating 24 readings as independent.
Common mistake: Testing matched measurements as independent discards the pairing and misrepresents the sampling structure, even when both raw groups look approximately normal.
Official syllabus reference: [1]
30. Variance tests and robustness boundaries
For independent normal observations, a one-sample variance test uses (n − 1)s²/σ0² and a chi-square reference distribution with n − 1 degrees of freedom. A classical two-variance F-test compares sample variances but is sensitive to nonnormality. Levene-type procedures compare dispersion using deviations from group centers; median-centered versions offer greater robustness in many settings. A nonsignificant variance test does not prove equality, especially with limited data. Choose the dispersion method according to assumptions and the intended decision, and distinguish a variance comparison from a comparison of means.
Worked example: A normal-process sample of sixteen observations has standard deviation three units, and the reference standard deviation is two units. The variance statistic is 15 × 9/4 = 33.75, evaluated against the appropriate chi-square tail or tails for fifteen degrees of freedom.
Common mistake: Substituting standard deviations directly into a variance-ratio formula changes the statistic and produces an incorrect comparison with its reference distribution.
Official syllabus reference: [1]
31. One-way ANOVA, contrasts and sample-size planning
One-way ANOVA compares group means by dividing between-group mean square by within-group mean square. Under independent normal errors with common variance, the resulting F statistic tests whether all population means are equal. Rejection establishes that at least one differs, without identifying specific pairs. Planned contrasts or suitable multiple-comparison procedures answer targeted follow-up questions. Check residual shape, variance behavior and design independence. Planning requires a meaningful difference, error estimate, alpha, power and group structure; unequal variances may motivate Welch's ANOVA. Statistical significance alone does not select an operationally superior setting.
Worked example: Four adhesive formulations are tested with eight independently randomized specimens each. A significant ANOVA prompts prespecified comparisons against the current formulation, using an appropriate multiplicity approach, rather than claiming that every formulation differs from every other one.
Common mistake: Running many unadjusted pairwise tests after ANOVA can inflate the familywise false-positive risk and overstate which groups differ.
Official syllabus reference: [1]
32. Mann–Whitney and Kruskal–Wallis rank comparisons
Mann–Whitney compares two independent groups through their combined ranks; Kruskal–Wallis extends a rank-based comparison to several independent groups. They avoid requiring normal raw data, but they still require a suitable independent-sampling structure. Their general interpretation concerns distributional or rank differences. Interpreting a result specifically as a median difference needs additional shape and location-shift conditions. Unequal spreads can influence ranks even when medians coincide. A significant Kruskal–Wallis result requires appropriate follow-up comparisons to identify groups, with attention to multiplicity. Neither procedure is the default for paired observations.
Worked example: Three independent supplier samples have heavily skewed delivery times. The analyst uses Kruskal–Wallis to compare their rank patterns, then checks distribution shapes before describing the result as evidence about medians rather than broader distribution differences.
Common mistake: Calling every significant rank test a test of medians ignores how unequal spread or shape can affect the rank distributions.
Official syllabus reference: [1]
33. Mood's median test and information loss
Mood's median test compares independent groups using counts above and below a pooled median. This makes the procedure relatively insensitive to extreme observation magnitudes, but discards much of the information contained in the data. Its contingency-table implementation needs appropriate count adequacy and a stated rule for observations equal to the pooled median. It is useful when the median is the explicit target and shape differences make simple rank-location interpretations questionable. Robustness does not guarantee strong power; other defensible methods may detect relevant differences more efficiently under more restrictive assumptions.
Worked example: Two production lines have durations with different tail behavior and occasional extreme stoppages. If management's explicit question concerns median duration, Mood's test classifies observations around the pooled median, using a documented tie rule before comparing the resulting group counts.
Common mistake: Assuming a robust median test is always more powerful mistakes resistance to extreme values for efficient use of all available sample information.
Official syllabus reference: [1]
34. Friedman testing for blocked rank comparisons
The Friedman test compares several treatments observed within matched blocks by ranking treatments separately inside each block. Blocking removes between-block level differences from the treatment comparison. Typical examples include several methods tested on the same product types or several conditions evaluated by the same participants. Blocks should be independent, and the design must support the within-block comparison; standard Friedman analysis expects a complete block structure. Rejection indicates a treatment-related rank difference, followed by appropriate paired comparisons when needed. It does not estimate the size of a difference in the original measurement units.
Worked example: Six material batches are each processed using three cleaning methods. Because batches differ in initial contamination, the analyst ranks the three outcomes within each batch and applies Friedman rather than treating all eighteen observations as unrelated measurements.
Common mistake: Using Kruskal–Wallis for a complete matched-block experiment ignores the blocking structure and treats dependent within-block observations as independent groups.
Official syllabus reference: [1]
35. Sign and Wilcoxon signed-rank procedures
For one-sample or paired location questions, the sign test uses only whether observations or differences lie above or below the hypothesized center. The signed-rank test also ranks absolute differences and attaches their signs, retaining magnitude-order information. Its usual location interpretation requires a symmetric distribution of differences. The sign test does not require that symmetry but often has less power when signed-rank assumptions hold. Both need appropriate independence between units or pairs. State how zero differences and tied magnitudes are treated, since these affect effective sample size and calculation details.
Worked example: A maintenance change is evaluated on the same twenty machines. Strongly asymmetric downtime differences make a sign-based comparison attractive if the question concerns the direction of change, while reasonably symmetric differences could support the more informative signed-rank approach.
Common mistake: Checking symmetry of the separate before and after samples misses the relevant assumption, which concerns the distribution of paired differences.
Official syllabus reference: [1]
36. One- and two-proportion inference
A one-proportion test compares a binary population probability with a reference value. A two-proportion test compares probabilities from independent groups. Large-sample procedures require suitable expected successes and failures; exact or other small-sample methods may be preferable when counts are sparse. For the usual null of equal proportions, a pooled estimate is commonly used in the test standard error, while confidence intervals may use different estimators. Paired binary outcomes require a matched analysis rather than an independent-proportions test. Report the difference and uncertainty alongside significance, using comparable definitions and exposure.
Worked example: Independent audits find 18 defective units among 300 before a change and 12 among 400 afterward. The observed reduction is three percentage points. The analyst evaluates uncertainty and sampling comparability before treating that difference as evidence of a sustained process improvement.
Common mistake: Applying an independent two-proportion test to the same units classified twice ignores pairing and can misstate the uncertainty of the change.
Official syllabus reference: [1]
37. Chi-square independence and cell contributions
Pearson's chi-square test assesses association between categorical variables using observed and expected cell counts. Under independence, each expected count equals its row total times column total divided by the grand total. Sum (O − E)²/E across all cells, with degrees of freedom (rows − 1)(columns − 1). Adequate expected counts and independent observations support the usual approximation. A significant result identifies association, while residuals help locate contributing patterns; neither proves causation. Sparse tables may require exact or other suitable methods, and collapsing categories needs substantive justification rather than p-value convenience.
Worked example: A two-by-three table classifies defect status across three shifts. The analyst calculates expected counts from both margins, includes contributions from conforming and defective cells, and examines which cells deviate most after finding evidence of an overall association.
Common mistake: Calculating chi-square from defective cells alone omits information in the complementary categories and produces a statistic with the wrong value.
Official syllabus reference: [1]
Improve Phase
38. Correlation, slope and simple linear regression
Pearson correlation measures the strength and direction of a linear association on a scale from −1 to +1. Simple linear regression models a conditional mean as β0 + β1X, with slope expressed in response units per predictor unit. Correlation is unitless and symmetric in the variables; regression assigns predictor and response roles. A strong relationship can reflect confounding or common trends rather than a causal mechanism. Inspect scatterplots for curvature and influential points, and avoid extrapolating beyond the supported range. A near-zero correlation does not exclude a strong nonlinear relationship.
Worked example: A fitted model predicts curing time as 80 − 0.4T within temperatures of 100–140°C. A ten-degree increase predicts four fewer minutes within that range, but the observational fit alone does not establish that changing temperature safely causes the reduction.
Common mistake: Interpreting correlation as the fraction of output caused by the predictor confuses association with causal attribution and misreads the statistic's scale.
Official syllabus reference: [1]
39. Multiple regression and conditional coefficients
Multiple linear regression estimates a response mean from several predictors. A coefficient describes the modeled change associated with one predictor while other included predictors are held fixed. This conditional interpretation can differ from a simple two-variable relationship. Multicollinearity makes individual coefficients unstable and increases their uncertainty, even when overall predictions remain reasonable. Include interactions when effects plausibly depend on other predictors, respecting model hierarchy when appropriate. Predictor inclusion does not automatically eliminate confounding, and coefficients from observational data remain association estimates unless the design and assumptions support stronger causal conclusions.
Worked example: A model includes order size and staffing level to predict dispatch time. The staffing coefficient describes differences at equal modeled order size, helping separate workload from staffing association, but unmeasured order complexity could still distort the apparent staffing effect.
Common mistake: Reading a multiple-regression coefficient as an unconditional effect ignores the other predictors held fixed and any interactions included in the fitted model.
Official syllabus reference: [1]
40. Polynomial models versus nonlinear regression
A model can be curved in a predictor while remaining linear in its coefficients. For example, β0 + β1X + β2X² is a linear regression model with a quadratic term. Nonlinear regression has parameters entering nonlinearly, such as an exponential-decay rate, and typically requires iterative estimation. Choose a functional form using process knowledge, experimental range and diagnostic evidence. Extra flexibility can overfit limited data. Scaling and centering predictors can improve numerical behavior and interpretation, but neither validates extrapolation. Physical constraints may favor a mechanistic nonlinear model over an unrestricted polynomial.
Worked example: A quadratic temperature model predicts a response turning point within the studied range. The analyst checks uncertainty and whether the point is a maximum or minimum before recommending confirmation runs, rather than treating the fitted vertex as an established operating optimum.
Common mistake: Calling every curved response nonlinear regression overlooks that polynomial models remain linear in their unknown coefficients and use ordinary linear-model machinery.
Official syllabus reference: [1]
41. Residual diagnostics and influential observations
Residuals are observed responses minus fitted responses. Plotting residuals against fitted values can reveal curvature or unequal variance; plotting them in time order can reveal dependence or drift. Probability plots help assess error-distribution assumptions relevant to inference. Leverage identifies unusual predictor combinations, while influence concerns how strongly an observation changes the fitted model. A high-leverage point is not automatically erroneous, and a large residual is not automatically influential. Investigate data quality and process circumstances before excluding observations. High R² cannot substitute for a model whose residual behavior supports the intended use.
Worked example: A regression has R² of 0.94, but residuals alternate between long positive and negative runs in production order. The analyst investigates temperature cycles and time dependence because the apparent fit does not establish valid independent-error inference.
Common mistake: Deleting the largest residual solely to improve R² can conceal genuine process behavior and make the model's reported performance unrealistically favorable.
Official syllabus reference: [1]
42. Mean-response confidence versus individual prediction
A confidence interval for a regression mean response quantifies uncertainty in the estimated conditional mean at a chosen predictor setting. A prediction interval for one future observation includes that estimation uncertainty plus individual residual variation, making it wider under the usual model. Both typically widen away from the center of the observed predictor configuration because estimation leverage increases. Neither is a process tolerance interval or a guarantee about all future observations. Interpretation requires the future case to match the modeled conditions and the regression assumptions to remain credible.
Worked example: At a supported setting, a model gives a mean-response interval of 24–26 minutes and an individual prediction interval of 17–33 minutes. The first informs average capacity planning, while the second better describes uncertainty for a single customer's processing time.
Common mistake: Using the narrow mean-response interval to promise an individual delivery time ignores the residual variation among otherwise comparable individual transactions.
Official syllabus reference: [1]
43. Box–Cox transformation and back-transformation
The Box–Cox family transforms a positive response using (Y^λ − 1)/λ when λ is nonzero and log(Y) when λ is zero. It can help stabilize variance and improve linear-model residual behavior. Select a transformation for the model's needs, then reassess diagnostics; transformed raw-data normality alone is insufficient. Standard Box–Cox requires positive values, and adding a constant changes interpretation. Back-transforming a fitted transformed mean does not generally produce the original-scale arithmetic mean because nonlinear transformations change expectations. Report the scale and use suitable bias correction when the original-scale mean is required.
Worked example: A log-response model predicts log duration as 3.0 at a setting. Exponentiating gives about 20.1 time units, representing a median-type prediction under a lognormal error model; estimating the arithmetic mean additionally requires accounting for residual variance.
Common mistake: Exponentiating a fitted log mean and labeling it the arithmetic mean ignores the difference between transforming an expectation and taking an expectation.
Official syllabus reference: [1]
44. DOE randomization, replication and blocking
Designed experiments deliberately vary factors to estimate effects under controlled comparisons. Randomization protects against systematic alignment with unknown time-related influences. Replication repeats independent experimental units to estimate experimental error; repeated readings of the same unit instead estimate measurement variation. Blocking groups runs by a known nuisance source, reducing its interference with treatment comparisons. Define the experimental unit, feasible factor ranges, response and objective before choosing a design. Restricted randomization may be necessary, but it changes the error structure and analysis. A well-organized run sheet does not replace these design principles.
Worked example: An adhesive experiment spans two material batches. Each batch contains a randomized set of treatment combinations, and independently prepared specimens provide replication. Taking three strength readings from one specimen is recorded as repeated measurement rather than three independent treatment replicates.
Common mistake: Treating repeated measurements on one experimental unit as independent replication understates experimental uncertainty and creates more apparent information than the design contains.
Official syllabus reference: [1]
45. Full factorial effects and interaction contrasts
A two-level full factorial with k factors contains 2^k treatment combinations before replication. It estimates main effects and interactions without their fractional-design aliasing. A main effect is the average high-minus-low response difference across other factors. An interaction means one factor's effect changes with another factor's level. With −1/+1 coding, a regression main-effect coefficient is half the corresponding factorial effect in the usual balanced model. When interaction matters, interpret conditional effects instead of recommending a setting from an averaged main effect alone. Randomization and replication remain essential design considerations.
Worked example: Increasing mixing speed raises yield by eight points at low temperature but lowers yield by two points at high temperature. The ten-point difference between those speed effects indicates interaction, so one overall speed recommendation would conceal an important operating dependence.
Common mistake: Choosing every factor's individually favorable main-effect level can fail when interactions make those effects change across the combined operating settings.
Official syllabus reference: [1]
46. Orthogonality, center points and curvature limits
A balanced two-level factorial allocates equal replication across treatment combinations. Its coded model columns are orthogonal for the supported effects, simplifying estimation because contrasts do not overlap. Missing runs or unequal replication can disrupt that property. Center points compare response near the design center with the factorial-corner pattern and, when independently replicated, help estimate pure error. They can detect overall curvature but do not separately estimate every quadratic coefficient. Fitting a full quadratic response surface generally needs additional factor levels, such as axial points. Lack-of-fit assessment also requires an appropriate error estimate.
Worked example: A two-factor corner design predicts the center response as 60 units, while replicated center runs average 68. The discrepancy suggests curvature, prompting an augmented response-surface design rather than an unsupported attempt to estimate separate squared effects from corners and center alone.
Common mistake: Assuming center points alone identify each factor's quadratic effect overlooks that the individual squared-factor terms are not separately resolved by that design.
Official syllabus reference: [1]
47. Fractional factorial resolution and alias structure
Fractional factorial designs reduce runs by accepting confounding among effects. A generator defines the fraction and produces a defining relation used to derive aliases by multiplication. Resolution is the length of the shortest nonidentity word: resolution III can alias main effects with two-factor interactions; resolution IV separates those classes but may alias two-factor interactions together; resolution V separates main and two-factor effects from each other, subject to higher-order aliases. Interpretation therefore depends on explicit assumptions about negligible effects. Choose augmentation from the actual alias structure rather than assuming every foldover resolves every ambiguity.
Worked example: For D = ABC, the defining relation is I = ABCD. Multiplying by AB gives AB = CD, so an apparent AB signal might be CD. Running the complementary D = −ABC fraction separates these effects by completing the full design.
Common mistake: Repeating the same fractional treatment combinations improves precision but leaves their structural aliasing unchanged, even when the repeat order is randomized.
Official syllabus reference: [1]
Control Phase
48. Sustaining 5S through visible operating standards
Sustained 5S converts an initial workplace cleanup into a repeatable operating condition. Define what belongs in each location, required quantities, cleaning and inspection responsibilities, and visible evidence of abnormal conditions. Standardization makes correct arrangements recognizable; self-discipline maintains them through routine practice and meaningful checks. Audits should identify deviations linked to safety, quality or flow, rather than merely producing attractive scores. Updating standards after process changes prevents obsolete layouts from becoming barriers. Ownership belongs with normal operations, with escalation for recurring causes that local housekeeping alone cannot resolve.
Worked example: A test bench uses labeled instrument locations and a start-of-shift check for missing leads and damaged connectors. A recurring missing-lead pattern triggers a replenishment change, demonstrating how the visual standard supports reliable work rather than a cosmetic inspection.
Common mistake: A high housekeeping score can coexist with poor process control when the audit rewards appearance but ignores missing items that affect correct operation.
Official syllabus reference: [1]
49. Kanban sizing and pull replenishment
Kanban authorizes production or movement in response to consumption, linking replenishment to a defined container or quantity. A common sizing relationship is cards = demand rate × replenishment lead time × (1 + safety factor) / container capacity, rounded up, with consistent time units. The safety factor is a scenario or policy assumption, not a universal constant. Lead-time variability, demand changes and quality losses affect required coverage. Kanban constrains work in process only when authorization rules are respected. Excess cards can hide instability, while insufficient coverage can cause avoidable shortages.
Worked example: Demand is 40 units hourly, replenishment takes three hours, containers hold twenty units, and the stated safety allowance is 10%. Required cards are 40 × 3 × 1.10/20 = 6.6, rounded up to seven authorized containers.
Common mistake: Adding emergency containers without reviewing card authorization increases hidden inventory and weakens the work-in-process control the pull system was designed to provide.
Official syllabus reference: [1]
50. Poka-yoke prevention versus detection
Poka-yoke changes work or equipment so that common errors are prevented or immediately detected before they propagate. Prevention mechanisms constrain incorrect assembly or entry; detection mechanisms identify a discrepancy and trigger a defined response. A warning is weaker than an effective interlock when operators can routinely bypass it. Select the mechanism from the actual error pathway and verify both correct acceptance and incorrect-case rejection. Mistake-proofing does not eliminate every possible failure mode, so its scope, maintenance needs and bypass behavior require explicit attention during handover.
Worked example: A connector fixture accepts a component only in the correct orientation, preventing a known reversal error. A separate sensor confirms full seating before release. The team challenges both mechanisms with incorrect and partially assembled examples during verification.
Common mistake: Calling a reminder poster mistake-proofing overstates its strength when the process still permits the same error without immediate detection or constraint.
Official syllabus reference: [1]
51. SPC anatomy and rational subgroup design
A control chart displays a time-ordered statistic, center line and limits derived from expected common-cause variation. Specifications express requirements and are separate from those limits. Rational subgroups collect observations so within-subgroup variation reflects short-term common causes, while between-subgroup comparisons can reveal meaningful changes. Sampling frequency should be capable of detecting shifts before unacceptable exposure accumulates. Choose signal rules before reviewing the chart and recognize that adding many rules changes false-alarm behavior. Establish baseline limits from a credible process period, investigate special causes and document any justified recalculation.
Worked example: A machining team samples five consecutive pieces every hour rather than mixing one piece from each shift into a subgroup. This preserves short-term variation within subgroups and makes shift-related changes visible between the hourly subgroup averages.
Common mistake: Mixing different process states inside each subgroup can inflate within-subgroup spread and make genuine changes between states harder to detect.
Official syllabus reference: [1]
52. I-MR estimation from successive observations
An I-MR chart monitors continuous measurements collected individually. The moving range of two is the absolute difference between successive observations, and its average estimates short-term spread using σ-hat = MR-bar/d2. Individual limits are commonly mean ± 3σ-hat; the moving-range chart separately monitors changing dispersion using appropriate constants. Inspect both charts before interpreting stability. Serial correlation can distort the moving-range estimate and signal behavior, so observations must represent a defensible sampling structure. Do not treat control limits as product tolerances or automatically recalculate them after an unexplained unusual point.
Worked example: Individual readings have mean 100 units and average moving range 2.256 units. With d2 = 1.128, estimated standard deviation is two units, giving individual limits of 94 and 106. A reading of 107 prompts the defined investigation.
Common mistake: Using the signed average of successive differences instead of absolute moving ranges can cancel changes and drastically understate estimated process variation.
Official syllabus reference: [1]
53. X-bar–R charts and range-first interpretation
An X-bar–R pair monitors subgroup means and ranges for continuous data, commonly with relatively small, consistent subgroup sizes. The mean-chart limits are grand mean ± A2R-bar; the range-chart limits are D3R-bar and D4R-bar, using constants matched to subgroup size. Examine the range chart first because changing within-subgroup dispersion undermines the mean chart's baseline estimate. Rational subgroups are essential: large differences deliberately mixed within a subgroup can mask between-subgroup movement. Variable subgroup sizes require suitable adjustments rather than silently applying one fixed set of constants.
Worked example: For subgroups of five, a scenario supplies A2 = 0.577, grand mean 20 and average range two. Mean-chart limits are 20 ± 1.154. The analyst first confirms that the range chart supports the variation estimate used in those limits.
Common mistake: Interpreting the mean chart while ignoring an unstable range chart overlooks that its control limits rely on the within-subgroup variation estimate.
Official syllabus reference: [1]
54. X-bar–S charts and standard-deviation estimation
An X-bar–S chart pair uses subgroup standard deviations instead of ranges to monitor within-subgroup dispersion. It is often preferred for larger subgroup sizes because the standard deviation uses more of the observations' information. Common limits use grand mean ± A3S-bar for subgroup means and B3S-bar to B4S-bar for subgroup standard deviations, with constants appropriate to subgroup size. S-bar requires a bias adjustment when used to estimate the underlying process standard deviation. Analyze dispersion first, preserve rational subgrouping and accommodate changing subgroup sizes with suitable methods.
Worked example: A laboratory collects twenty consecutive measurements per subgroup. It uses an S chart to evaluate within-subgroup spread, then an X-bar chart for averages, selecting the supplied constants for twenty observations rather than borrowing constants from a five-piece range chart.
Common mistake: Switching from ranges to standard deviations while retaining X-bar–R constants gives limits that do not match the statistic's sampling behavior.
Official syllabus reference: [1]
55. p and np charts for nonconforming units
A p chart monitors the proportion of units classified nonconforming, with each unit counted once. Its usual limits are p-bar ± 3√[p-bar(1 − p-bar)/n], truncated to feasible proportions as needed. An np chart monitors the count nonconforming with a constant sample size, using center np-bar and corresponding binomial limits. Varying sample sizes change proportion uncertainty and therefore p-chart limits. Binary independence and a comparable classification process matter; clustering or overdispersion may invalidate simple binomial limits. Neither chart is intended to count multiple distinct defects within the same unit.
Worked example: Daily audits classify each shipment as complete or incomplete, with audit sizes ranging from eighty to two hundred. A p chart uses the daily denominator and varying limits, avoiding the misleading comparison that results from plotting incomplete-shipment counts alone.
Common mistake: Using one fixed set of count limits when inspected volume changes confuses increased exposure with a change in the probability of nonconformance.
Official syllabus reference: [1]
56. u charts for defects per exposure
A u chart monitors counted nonconformities per unit of exposure, allowing multiple defects within a unit. Estimate u-bar as total defects divided by total exposure, rather than averaging subgroup rates without appropriate weights. Standard Poisson-based limits for subgroup exposure n are u-bar ± 3√(u-bar/n), with a nonnegative lower limit. Exposure can represent units, area or another consistently defined opportunity basis. Varying defect clustering or opportunity quality can create overdispersion, so verify the model and denominator. A u chart differs from a p chart because its numerator counts defects, not defective units.
Worked example: Inspectors count surface blemishes on panels of different areas. They report blemishes per square meter and use each subgroup's inspected area in the limits, ensuring a larger inspected surface is treated as more exposure rather than automatically worse quality.
Common mistake: Taking an unweighted average of defect rates gives small-exposure and large-exposure subgroups equal influence and can misestimate the overall center line.
Official syllabus reference: [1]
57. CUSUM reference values and accumulated evidence
CUSUM charts accumulate small departures from a target, improving sensitivity to sustained shifts that may not create extreme individual points. A one-sided standardized form is C-plus,t = max(0, C-plus,t−1 + zt − k), signaling above a decision interval h. The reference value k discounts routine small deviations, while h controls the evidence needed to signal. A lower-side CUSUM tracks decreases. Parameter choice depends on the shift of interest and desired false-alarm behavior. Use an appropriate stable baseline and defined restart procedure; the cumulative statistic is not interpreted using ordinary individual-point limits.
Worked example: With k = 0.5 and initial cumulative value zero, standardized observations of 0.9, 1.0 and 0.8 produce cumulative values 0.4, 0.9 and 1.2. A scenario's specified decision interval determines whether and when this accumulation signals.
Common mistake: Adding standardized observations without subtracting the reference value changes the chart's intended sensitivity and does not implement the specified tabular CUSUM.
Official syllabus reference: [1]
58. EWMA weighting, memory and time-varying limits
An exponentially weighted moving average updates Zt = λXt + (1 − λ)Zt−1, giving recent observations more weight while retaining decaying historical influence. Smaller λ creates longer memory and can improve sensitivity to small sustained shifts, but also slows response to abrupt changes. Under independent observations with known stable variance, the EWMA variance approaches σ²λ/(2 − λ); startup variance additionally includes a time-dependent factor. Set the starting value and limits consistently with the design. Autocorrelation or a drifting baseline affects signal behavior, so smoothing alone does not establish valid process control.
Worked example: With λ = 0.2, starting value 50 and a new observation of 55, the EWMA becomes 51. If the next observation is also 55, it becomes 51.8, accumulating evidence without jumping immediately to either individual reading.
Common mistake: Comparing the smoothed statistic directly with raw-observation limits ignores the EWMA's different variance and can greatly reduce the intended sensitivity.
Official syllabus reference: [1]
59. Control plans and executable response plans
A control plan identifies critical outputs and inputs, measurement methods, sampling, acceptance or signal rules, records and accountable roles. Its response plan defines what happens when a rule is triggered: verify the measurement, contain potentially affected output, investigate the cause, correct it and authorize restart using specified evidence. Distinguish product nonconformance from statistical instability because either can occur without the other. Include escalation and traceability boundaries so containment is neither arbitrary nor incomplete. Routine common-cause variation should lead to system improvement when needed, rather than repeated local adjustments that increase variation.
Worked example: A temperature alarm triggers a defined check of sensor validity, isolation of material processed since the last verified acceptable condition, maintenance escalation and restart only after documented stable readings. The named shift lead records disposition and completion evidence.
Common mistake: Writing 'adjust as necessary' leaves operators without a common trigger or response and can encourage tampering with an otherwise stable process.
Official syllabus reference: [1]
60. Realized benefits, balancing metrics and sustainment
Sustainment verifies that improved performance and its economic value persist after project handover. Compare realized outcomes with the agreed baseline using consistent definitions, comparable demand and an appropriate treatment of changing product mix. Deduct recurring control and maintenance costs when reporting net benefits. Separate cash savings, cost avoidance, capacity and revenue contribution, and prevent overlap. Balancing metrics reveal displaced costs or quality losses, such as faster service achieved through more rework. The process owner maintains controls, while finance or another designated function validates the claimed benefit using an auditable reconciliation.
Worked example: A project claims fewer overtime hours, but demand also fell. The team adjusts its comparison for workload, checks that defect rates did not rise, subtracts new system support costs, and obtains finance validation before reporting recurring net cash savings.
Common mistake: Multiplying a favorable pilot result by annual volume without checking sustained performance, demand and recurring costs can substantially overstate realized benefits.
Official syllabus reference: [1]