Apply DMAIC to select worthwhile projects, build trustworthy measurements, distinguish plausible causes from evidence and sustain verified improvements. This guide combines twenty original multiple-choice questions with sixty worked concepts, including capability calculations, test selection, regression interpretation and control-chart decisions. Coverage follows the January 2026 ICGB Version 1.1 syllabus supplied for the scope current as of 2026-10-16, using the prescribed preview allocation. It addresses initial certification, rather than recertification or Black Belt extensions. This is a representative preview, not a full-length examination. Source citations establish syllabus relevance; the worked technical explanations apply established Lean Six Sigma knowledge.
What this guide covers
Twenty original ICGB practice questions and 60 focused concepts covering DMAIC, statistics and Lean controls.
Define Phase – Project Selection, Chartering and Scoping
1. DMAIC and the meanings of Six Sigma
Six Sigma can describe a statistical performance aspiration, an improvement methodology and an organizational system for managing improvement. Its modern development is associated with Motorola, while Lean draws strongly on the Toyota Production System. DMAIC structures work on an existing process: Define establishes the problem, Measure establishes trustworthy performance, Analyze tests explanations, Improve changes the process and Control sustains results. Each phase produces evidence needed by the next. A quoted sigma level is meaningful only with its opportunity definition, distribution assumptions and any stated shift convention; it is not automatically proof of excellence.
Worked example: A service team first defines late activation, validates timestamp data and tests causes before changing scheduling. Its Control handover includes monitoring and a response plan. Buying scheduling software at the outset would bypass the evidence-producing sequence that DMAIC provides.
Common mistake: Treating Six Sigma solely as a defect-rate label overlooks the project method and the management responsibilities needed to sustain improvement.
Official syllabus reference: [1]
2. Y = f(X): outputs, inputs and causal verification
The expression Y = f(X) frames a process output as a function of inputs and operating conditions. Y is the response to improve; candidate X variables are possible drivers. A process measure can serve as an intermediate Y at one level and an X for a downstream outcome. The expression encourages movement from managing symptoms to understanding mechanisms, but it does not establish that a correlated X causes Y. Teams generate hypotheses through process knowledge, then use suitable data and comparisons to identify which inputs actually explain meaningful variation.
Worked example: For adhesive curing, bond strength is Y, while temperature, cure time and mixture ratio are candidate X variables. A correlation between temperature and strength prompts investigation, but simultaneous changes in mixture ratio must be addressed before attributing the effect to temperature.
Common mistake: Calling every measured input a root cause confuses a candidate explanatory variable with a relationship supported by process evidence.
Official syllabus reference: [1]
3. VOC, VOB and VOE translated into CTQs
Voice of the customer identifies valued outcomes, voice of the business identifies organizational requirements and voice of the employee reveals practical operating needs. A CTQ translates a relevant need into a measurable characteristic with a clear definition and requirement. Distinguish the customer's output requirement from a proposed solution or internal activity measure. Conflicting voices should become explicit tradeoffs rather than hidden assumptions. A CTQ tree can move from a broad need through performance drivers to measurable characteristics, helping connect project metrics to the people and purposes the process serves.
Worked example: Customers request predictable delivery, finance wants fewer expedite charges and dispatchers want stable schedules. The team defines an on-time-delivery CTQ using the original agreed date, then uses expedite cost and schedule changes as supporting measures to reveal tradeoffs.
Common mistake: A request to install an application is a solution preference; it does not by itself define the customer outcome the application should improve.
Official syllabus reference: [1]
4. Sponsor, Green Belt and process-owner accountability
Improvement roles separate strategic authority, analytical work and ongoing process ownership. A sponsor or champion aligns the project with organizational priorities and helps remove barriers. A Green Belt applies improvement methods and coordinates project work within the organization's arrangements. Process owners remain accountable for the operating process and must accept sustainable controls. Subject-matter experts contribute practical knowledge, while more advanced belt roles can provide methodological support. Exact job boundaries vary, so a charter should make decision rights explicit. Successful project closure requires more than an analyst producing a statistically favorable result.
Worked example: A Green Belt validates a reduction in picking errors, the sponsor resolves a cross-department resource conflict and the warehouse owner accepts the revised work standard and chart response. Named ownership ensures the improvement continues after the project team disbands.
Common mistake: Assuming the project leader permanently owns the improved process leaves routine monitoring, maintenance and escalation responsibilities unclear after handover.
Official syllabus reference: [1]
5. Charter boundaries and solution-neutral problem statements
A project charter connects the business case, verified problem, measurable goal, scope, roles and expected timing. The problem statement describes a performance gap without claiming an untested cause or prescribing a solution. Scope defines process start and end points, relevant populations and exclusions so the team can complete meaningful work with available authority and resources. A goal should specify the metric, baseline, target and deadline when those are established. Boundaries can be revised through agreed governance, but uncontrolled expansion dilutes resources and makes results harder to attribute.
Worked example: A charter targets reducing correction-required domestic invoices from 12% to 6% by September, covering invoice creation through first submission. Collections and international tax handling are excluded. The wording leaves room for different causes rather than committing to new software.
Common mistake: Writing the preferred technology into the problem statement converts a root-cause project into implementation before the proposed solution has been justified.
Official syllabus reference: [1]
6. Cost of poor quality and failure-cost boundaries
Cost of poor quality captures resources lost because outputs or processes fail to meet requirements. Internal failure costs arise before delivery, such as scrap, rework and retesting. External failure costs arise after delivery, such as returns, warranty work and complaint handling. Prevention and appraisal are different quality-cost categories, although accounting conventions should be stated explicitly. Estimate failure costs using actual volumes, resource consumption and relevant rates, while avoiding double counting. Opportunity costs and released capacity can be valuable, but they should be distinguished from reductions in cash expenditure.
Worked example: A batch generates $2,000 of scrap and 50 rework hours valued at $30 each. Its measured failure cost is $3,500 before any external claims. If employees remain on payroll, eliminating the rework releases capacity without automatically saving $1,500 in cash.
Common mistake: Adding the full value of rejected goods to material scrap and all embedded labor can count the same loss more than once.
Official syllabus reference: [1]
7. Pareto analysis by frequency, cost or consequence
A Pareto chart ranks categories from largest to smallest and may add a cumulative percentage line. It helps concentrate attention where a chosen measure is most concentrated. The familiar 80:20 expression is a heuristic, not a required empirical split. Category definitions must avoid overlap, and the ranking changes when the measure changes from count to cost, delay or another consequence. A chart prioritizes investigation; it does not prove root causes. Rare but severe failures may require attention even when they appear small in a frequency-based ranking.
Worked example: A team records 60 labeling errors costing $5 each and 20 damage incidents costing $100 each. Labeling leads the frequency Pareto, but damage leads the cost Pareto at $2,000 versus $300. The project priority depends on the stated objective.
Common mistake: Assuming the tallest bar identifies a verified root cause mistakes a category of observed failures for an explanation of their mechanism.
Official syllabus reference: [1]
8. DPU and DPMO use different denominators
Defects per unit equals total defects divided by units examined. DPMO equals total defects divided by total applicable opportunities, multiplied by one million. A unit may contain multiple defects, so neither measure is automatically the proportion of defective units. Opportunity definitions must represent meaningful, consistently counted chances for failure. When opportunity counts vary by unit, add the actual opportunities rather than multiplying by an unsupported common count. Comparisons become misleading if teams redefine opportunities to improve the reported metric without improving the underlying customer experience.
Worked example: Inspectors find 45 defects in 300 forms, each with five applicable checks. DPU is 45 divided by 300, or 0.15. DPMO is 45 divided by 1,500, multiplied by one million, or 30,000; the defective-form count is still unknown.
Common mistake: Using defective units in the DPMO numerator undercounts failures when a single inspected unit contains more than one defect.
Official syllabus reference: [1]
9. First-time yield and rolled throughput yield
First-time yield measures the fraction of units passing a process step without repair, rework or repeat processing. Rolled throughput yield estimates the fraction passing every step first time by multiplying appropriate step yields. Use yields from a consistent process path and population; branching routes require explicit handling. Final yield after rework can look excellent while concealing a costly hidden factory of repeated work. RTY highlights the cumulative effect of several modest step losses, but multiplying unrelated yields from incompatible product mixes can create a misleading result.
Worked example: Three sequential steps have first-time yields of 0.95, 0.90 and 0.98 for the same routing. RTY is 0.8379, or 83.79%. Even if rework eventually makes every shipped unit conforming, about 16.21% fail at least one first pass.
Common mistake: Averaging step yields does not calculate the probability of passing the complete sequence without rework or repeated processing.
Official syllabus reference: [1]
10. Cycle time, lead time and takt time
Cycle time describes processing time or the interval between completed units, depending on the stated operational definition. Lead time is elapsed time from a defined request or entry point to completion and can include queues, transport and processing. Takt time is available production time divided by customer demand for that period. Takt expresses the demand pace, while observed cycle time describes process performance. Comparing them helps reveal capacity mismatches. Always state the timing boundaries and whether activities overlap; adding machine times does not necessarily equal the customer's elapsed experience.
Worked example: A cell has 420 available minutes and demand for 210 units, giving a takt time of two minutes per unit. Its observed completion interval is 2.4 minutes, indicating a capacity shortfall unless the work arrangement or available capacity changes.
Common mistake: Calling all elapsed customer waiting time cycle time can hide queues and make comparisons between differently defined process measures unreliable.
Official syllabus reference: [1]
11. Project selection, cash benefits and financial evaluation
Project selection weighs strategic relevance, customer impact, measurable opportunity, feasibility and the authority to change the process. Financial evaluation then compares incremental benefits with implementation and recurring costs. Simple payback divides initial investment by annual net cash benefit when that benefit is constant. Net present value discounts future cash flows using an explicitly chosen rate. Separate cash savings, cost avoidance, revenue effects and released capacity, and identify who will validate each benefit. Avoid claiming the same saving in several projects or treating an accounting valuation as cash without a realizable mechanism.
Worked example: An improvement costs $40,000 and avoids $18,000 annually while adding $2,000 of maintenance. Net annual cash benefit is $16,000, so simple payback is 2.5 years. Finance separately records 300 released hours because staffing expenditure will remain unchanged.
Common mistake: Selecting the largest advertised saving without checking feasibility and benefit realization can favor projects whose financial returns cannot actually be captured.
Official syllabus reference: [1]
12. The seven wastes and Lean flow
The detailed ICGB syllabus lists seven wastes: overproduction, correction, inventory, motion, overprocessing, conveyance and waiting. Overproduction creates output before demand; inventory stores accumulated work; conveyance moves material; motion is unnecessary human movement. Correction repairs failures, overprocessing performs unnecessary work and waiting consumes elapsed time without progress. Lean improves customer value and flow, while Six Sigma reduces variation and defects. Some activities are currently necessary without adding customer value, so classification should distinguish elimination opportunities from required controls. Additional waste frameworks should not replace the detailed syllabus's seven-category coverage.
Worked example: A clerk walks to a shared printer, carries forms between desks and waits for a signature. These illustrate motion, conveyance and waiting respectively. Printing extra copies before demand adds overproduction, even when each copy is produced quickly and without defects.
Common mistake: Treating every required control as immediately removable waste ignores why the control exists and whether an alternative can meet the requirement.
Measure Phase – Process Definition, MSA and Capability
13. SIPOC establishes process interfaces
SIPOC identifies suppliers, inputs, a high-level process, outputs and customers. It establishes the process boundary and interfaces before detailed mapping. Suppliers can be internal, and customers include recipients who use the output rather than only paying purchasers. Inputs should be distinguished from the organizations supplying them, while outputs should be distinguished from activities. A SIPOC is deliberately concise: it supports agreement on what process is being studied and whose requirements matter. It does not show enough timing or decision detail to diagnose every source of waste or variation.
Worked example: For access provisioning, an employee's manager supplies an approved request, IT performs validation and setup, and the employee receives working access. The manager is a supplier, the approved request is an input and working access is the output, not the setup activity.
Common mistake: Filling a SIPOC with dozens of detailed tasks obscures the high-level interfaces it is meant to clarify before deeper mapping.
Official syllabus reference: [1]
14. Process maps reveal decisions, loops and handoffs
A process map represents the sequence of work, including decisions, branches, rework loops and handoffs. A swimlane map adds responsibility by locating activities within organizational lanes. Mapping the actual process requires observation and operator input because written procedures may omit informal workarounds. Clear start and end points prevent scope drift, while explicit decision criteria reveal inconsistent routing. Maps generate hypotheses about failure points and delays; they do not quantify their importance without data. Separate the current-state map from a proposed future state so improvement claims remain traceable.
Worked example: An expense map shows requests returning from finance to employees for missing receipts, then re-entering the same queue. A swimlane view reveals two department handoffs. Counting loop frequency and waiting time tests whether this route materially drives reimbursement delays.
Common mistake: Mapping only the approved procedure can miss the actual rework and informal routing responsible for much of the observed performance loss.
Official syllabus reference: [1]
15. Value stream maps separate work from delay
A value stream map connects material or service flow with the information that authorizes work. It typically adds process times, queues, inventory and other data to expose system-level delay. Value-added time should be distinguished from total lead time using explicit criteria and consistent boundaries. A fast local operation can still contribute to poor overall flow if it produces ahead of demand or builds a downstream queue. Future-state mapping therefore addresses the whole stream, including pull signals and workload relationships, rather than maximizing the apparent efficiency of each isolated step.
Worked example: An application requires twelve minutes of review but waits four hours before review and two hours afterward. With no overlap, its lead time is 372 minutes. The map directs attention toward queue policies rather than assuming faster review is the largest opportunity.
Common mistake: Adding process times alone and calling the sum lead time excludes the queues that often dominate the customer's elapsed experience.
Official syllabus reference: [1]
16. Fishbone diagrams organize testable causes
A cause-and-effect or fishbone diagram organizes possible explanations for a clearly defined effect. Common categories such as methods, machines, materials, people, measurement and environment prompt broad exploration, but categories should fit the process. Branches should become specific, testable statements rather than vague labels such as human error. The diagram records hypotheses, not verified causes. Subsequent measurement, stratification and testing determine which proposed mechanisms matter. Including measurement-system causes is important because apparent output variation can come from how the response is recorded rather than the process itself.
Worked example: For variable fill weight, a team lists nozzle wear, product temperature and scale drift. It converts these into checks of nozzle age, measured viscosity and reference-weight readings, allowing each branch to be investigated rather than simply voting for the most familiar explanation.
Common mistake: A heavily populated fishbone demonstrates breadth of brainstorming, but the number of branches does not establish that any proposed cause is true.
Official syllabus reference: [1]
17. X–Y matrices prioritize input investigation
An X–Y diagram or cause-and-effect matrix connects process inputs with important outputs. Outputs receive importance weights, and input-output relationships receive consistently defined scores. Summing weighted relationships gives a structured way to prioritize candidate inputs for investigation. The method combines process knowledge with customer priorities, making assumptions visible. Scores are judgments rather than statistical effect estimates, so a high-ranked input still needs verification. Avoid false precision from elaborate arithmetic when the underlying ratings are uncertain, and review rankings when important outputs or process knowledge change.
Worked example: A team weights leakage at five and appearance at two. An input with relationship scores of nine and one totals 47; another scoring three and nine totals 33. The first input receives investigation priority, but the matrix does not prove its causal effect.
Common mistake: Interpreting a weighted relationship score as a measured regression coefficient gives judgment-based prioritization more statistical authority than it has.
Official syllabus reference: [1]
18. FMEA links failure modes, effects and controls
FMEA anticipates how a process can fail, what each failure causes and why it might occur. Severity describes the consequence, occurrence describes the cause's frequency and detection describes the ability of current controls to detect the problem under the chosen scale. The traditional RPN multiplies these ratings, but equal products can conceal very different risks. Prioritization should examine the full profile, especially severe effects. Actions should target prevention, consequence reduction or detection as appropriate, with ownership and reassessment. A lower detection rating does not mean the failure itself happens less often.
Worked example: A missing seal and a cosmetic scuff both receive RPN 90 through different rating combinations. The team examines the missing seal's leakage consequence first, introduces an assembly interlock and reassesses occurrence and detection after verifying the new control's actual performance.
Common mistake: Reducing a detection score on paper without validating the control improves the worksheet while leaving the actual failure risk unchanged.
Official syllabus reference: [1]
19. Data types and operational definitions
Data type determines which summaries and methods are appropriate. Continuous measurements represent quantities such as length or time; discrete counts represent events or items. Attribute data may classify units into categories, including conforming and defective. Distinguish a defect count from a count of defective units because one unit can contain several defects. An operational definition specifies what is measured, the unit, decision rules, timing and handling of ambiguous cases. Numeric labels do not turn categories into continuous measurements. Reliable analysis begins with a consistent definition applied by all collectors.
Worked example: A team defines a late shipment as departure after the original agreed dispatch timestamp. It records departure delay in minutes and a separate late/not-late classification. Revising the promise after departure is prohibited from changing the baseline classification, preserving a consistent metric.
Common mistake: Assigning category codes one, two and three does not justify interpreting their average as a physical quantity with equal numerical intervals.
Official syllabus reference: [1]
20. Mean, median and weighted averages
The mean uses every numerical observation and is sensitive to extreme values. The median identifies the middle ordered value and is more resistant to skew and outliers. Neither is universally superior: the relevant center depends on the business question and distribution. A pooled mean requires weighting subgroup means by their observation counts when those counts differ. Reporting a center without spread can conceal very different process experiences. For skewed service times, both mean resource demand and median typical experience may be informative, while upper quantiles can reveal long delays.
Worked example: One branch handles ten requests averaging eight minutes; another handles ninety averaging four minutes. The pooled mean is (10 × 8 + 90 × 4) divided by 100, or 4.4 minutes. Averaging the two branch means would incorrectly give six minutes.
Common mistake: Choosing the median merely because it is smaller can hide the total workload created by a process with occasional very long cases.
Official syllabus reference: [1]
21. Variance, standard deviation and robust spread
Sample variance summarizes squared deviations from the sample mean, using n − 1 in the denominator for the usual unbiased population-variance estimator. Standard deviation is its square root and returns spread to the original measurement units. The range uses only extremes and is sensitive to sample size; the interquartile range summarizes the middle half and resists extreme observations. Independent variance components add, whereas standard deviations generally do not. Spread measures describe different features, so choose them with the distribution, subgroup structure and intended calculation in mind.
Worked example: Independent process and measurement components have standard deviations of three and four micrometres. Their variances are nine and sixteen, giving total variance twenty-five and total standard deviation five micrometres. Adding the standard deviations would incorrectly produce seven micrometres.
Common mistake: Treating variance as though it has the original measurement units can produce incorrect comparisons and dimensionally inconsistent calculations.
Official syllabus reference: [1]
22. Normal distributions and graphical diagnosis
A normal distribution is symmetric and described by its mean and standard deviation. Standardizing with z = (x − μ)/σ expresses distance from the mean in standard-deviation units. Histograms show shape, probability plots reveal departures from a reference distribution and time plots expose changes that an overall histogram can hide. A normality-test p-value should be considered alongside graphics and sample size: small samples can miss departures, while large samples can flag minor ones. Mixing stable subpopulations can create nonnormal pooled data even when each subgroup is approximately normal.
Worked example: Measurements from two machines form a double-peaked histogram. Separate plots show each machine is roughly normal around a different mean. Fitting one normal curve to the pooled measurements would obscure the machine difference and could misrepresent tail probabilities and capability.
Common mistake: A nonsignificant normality test does not prove normality, especially when the sample is too small to reveal meaningful departures.
Official syllabus reference: [1]
23. Accuracy and precision are separate properties
Precision concerns agreement among repeated measurements under specified conditions. Accuracy concerns closeness to the reference or true value and is affected by systematic and random error. An instrument can repeat tightly while consistently reading high, or have an average near the reference while individual readings scatter widely. Calibration may address an offset but does not automatically reduce repeatability variation or operator effects. Measurement resolution must also be fine enough for the intended discrimination. Evaluate performance against the actual measurement task rather than assuming a display with many digits is adequate.
Worked example: A balance repeatedly reads approximately 102 grams for a certified 100-gram reference. The narrow spread supports precision, while the two-gram offset indicates bias. Correcting that offset would not establish performance at other loads or prove consistent results across different operators.
Common mistake: Repeatedly obtaining the same value demonstrates consistency under those conditions, but it does not establish agreement with a trustworthy reference.
Official syllabus reference: [1]
24. Bias, linearity and stability diagnose different errors
Bias is the difference between the mean measured value and a reference value at a specified level. Linearity assesses whether bias changes across the measurement range. Stability assesses whether measurement performance changes over time. These dimensions require different study structures: repeated readings at one reference estimate local bias, multiple references explore linearity and repeated checks over time explore stability. A correction established at one value may be inadequate elsewhere. Reference uncertainty, environmental conditions and the intended operating range should be considered when interpreting apparent changes in the measurement system.
Worked example: A thermometer reads one degree high at 20°C and four degrees high at 80°C, suggesting changing bias across range. Daily checks at 20°C then reveal a gradual rise, indicating a separate stability concern that a one-time offset correction would miss.
Common mistake: A successful calibration at one reference level does not establish linearity across the range or stability throughout future operation.
Official syllabus reference: [1]
25. Crossed gage R&R separates measurement variation
A crossed gage R&R study has multiple appraisers measure the same parts repeatedly, usually in randomized blinded order. Repeatability captures variation under repeated measurement conditions, while reproducibility captures appraiser-related variation and potentially appraiser-by-part interaction. Parts should represent the intended process range. Independent variance components combine to estimate measurement-system variance. Percentage contribution uses variance ratios; percentage study variation uses standard-deviation ratios. Percentage of tolerance uses a specification-based denominator and answers another question. Interpretation must match the metric and intended decision rather than applying an unexplained universal acceptance cutoff.
Worked example: A study estimates gage variance of four and part variance of ninety-six square micrometres. The gage contributes 4% of total variance, but its study-variation percentage is the square root of 0.04, or 20%. Both statements describe the same components differently.
Common mistake: Comparing a variance-contribution percentage directly with a standard-deviation-based criterion mixes different metrics and can reverse the measurement-system judgment.
Official syllabus reference: [1]
26. Attribute agreement and false classification
Attribute MSA evaluates classification consistency and correctness. Within-appraiser agreement compares repeated judgments by one appraiser; between-appraiser agreement compares people; reference agreement compares judgments with a defensible classification. Overall agreement can conceal category-specific weaknesses when one class dominates. Examine false acceptance and false rejection separately, and include representative borderline cases. Chance-adjusted agreement statistics such as kappa add information but are affected by prevalence and marginal distributions. A deliberately enriched challenge sample is useful diagnostically, yet its overall error fraction is not automatically the production misclassification rate.
Worked example: An inspector accepts ninety-five conforming items correctly but misses four of five defective items. Overall agreement is 96%, while defect detection is only 20%. Repeated blinded classification of additional defective and borderline items reveals the weakness more clearly than another mostly conforming sample.
Common mistake: High overall agreement does not establish reliable defect detection when the study contains very few genuinely defective or borderline items.
Official syllabus reference: [1]
27. Process stability precedes capability prediction
A stable process exhibits a consistent pattern of common-cause variation under its defined operating conditions. Special causes introduce identifiable departures such as shifts, trends or unusual points. Time-ordered monitoring distinguishes these patterns in ways an overall histogram cannot. Stability does not mean the output meets customer specifications, and conformance does not prove stability. Capability prediction assumes a sufficiently consistent process distribution, so unresolved special causes undermine forward-looking interpretation. Investigate signals using process evidence, avoid automatic adjustment to every ordinary fluctuation and reassess the baseline after justified process changes.
Worked example: Every sampled diameter falls within specification, but the chart shows a sustained upward shift after tool replacement. The process is conforming in this sample yet unstable. Resolving the replacement effect is necessary before projecting future nonconformance from a single fitted distribution.
Common mistake: Using specification limits as evidence of statistical stability confuses customer requirements with the process's time-dependent pattern of variation.
Official syllabus reference: [1]
28. Cp and Cpk distinguish spread from centering
For suitable stable approximately normal data, Cp compares specification width with six within-process standard deviations: Cp = (USL − LSL)/(6σwithin). Cpk uses the smaller standardized distance from the mean to either specification limit. Thus Cp describes potential spread fit, while Cpk also penalizes off-centering. Cpk cannot exceed Cp when both are calculated consistently. A large gap suggests centering is limiting performance. Neither index supplies a universal acceptance decision without an agreed requirement, and neither justifies normal-tail predictions when distributional or stability assumptions are inappropriate.
Worked example: With limits 20 and 26, mean 25 and within standard deviation 0.5, Cp equals 2.0. Cpk is the smaller of 1 divided by 1.5 and 5 divided by 1.5, or about 0.67, revealing proximity to the upper limit.
Common mistake: A high Cp alone does not establish low nonconformance when the process mean sits close to one specification boundary.
Official syllabus reference: [1]
29. Pp and Ppk use overall variation
Pp and Ppk parallel Cp and Cpk but use the overall sample standard deviation rather than an estimate of within-subgroup variation. They summarize performance across the period represented by the sample. Differences between within and overall estimates can reveal shifts, drift or subgroup effects that deserve investigation. The distinction depends on how subgroups and variation estimators are defined; it is not automatically a fixed short-term versus long-term calendar distinction. On an unstable process, an overall index describes collected history but should not be promoted as a reliable prediction of future performance.
Worked example: A process has specification width twelve, within standard deviation one and overall standard deviation two. Cp is 2.0 while Pp is 1.0. The discrepancy prompts a review of between-shift changes rather than a claim that one calculation must be erroneous.
Common mistake: Calling Ppk a guaranteed future capability measure ignores that its overall spread may combine several changing process states.
Official syllabus reference: [1]
30. Nonnormal capability needs an appropriate distribution
Skewed or otherwise nonnormal data can make normal-based tail estimates misleading. First investigate whether the shape arises from mixed populations, instability, recording limits or an inherently nonnormal mechanism. For a stable homogeneous process, suitable alternatives include a justified distributional model, a valid transformation or percentile-based capability analysis. A transformation must also be applied consistently to specification limits, and results should remain interpretable in original units. Distribution choice affects predicted tail risk, so graphical fit and practical process knowledge matter, especially where few observations occur near the specification boundary.
Worked example: Stable repair times are strongly right-skewed. A normal fit assigns impossible negative durations and understates the long upper tail. The team compares a positive-valued model with probability plots and reports estimated exceedance of the stated service limit, including uncertainty.
Common mistake: Transforming only the measurements while leaving specification limits on their original scale creates a capability comparison between incompatible quantities.
Official syllabus reference: [1]
31. Discrete capability separates defectives from defects
Attribute capability can summarize the proportion of units that are defective or the rate of defects per unit or opportunity. A binomial framework suits binary unit classifications under its assumptions; a Poisson framework can suit defect counts over defined exposure when its assumptions are reasonable. Several defects can occur on one unit, so these frameworks are not interchangeable. Report the denominator, exposure and uncertainty, especially for small counts. Zero observed defects does not imply a zero underlying defect probability, and changing opportunity definitions can invalidate comparisons.
Worked example: Among 500 inspected assemblies, twenty are defective and contain thirty separate defects. The defective proportion is 4%, while defects per assembly is 0.06. An improvement that reduces multiple defects on already defective units can lower DPU without changing the defective proportion.
Common mistake: Reporting zero defects from a limited sample as perfect capability ignores sampling uncertainty and the possibility of rare failures.
Official syllabus reference: [1]
Analyze Phase – Root Cause Analysis and Hypothesis Testing
32. Multi-vari analysis locates variation families
Multi-vari analysis displays variation across meaningful nested or crossed factors, such as positions within a part, parts within a period and periods over time. Its purpose is to locate dominant patterns before committing to a detailed causal model. Sampling must preserve these relationships; randomly pooling all readings destroys the structure. Distribution classes should also match the response mechanism: continuous measurements, binary outcomes and event counts behave differently. Large apparent factor differences can suggest where to investigate, but observational patterns alone do not eliminate confounding or establish the physical cause.
Worked example: A team measures three positions on each of four tubes during every shift. Position differences dominate within each tube, while shift averages are similar. Investigation therefore focuses on the forming geometry before attributing the variation to a particular shift's operators.
Common mistake: A pooled histogram can show total spread while concealing whether variation arises within units, between units or across time periods.
Official syllabus reference: [1]
33. Sampling design controls the scope of inference
Statistical inference extends sample evidence to a defined population using assumptions about selection and dependence. Simple random sampling gives eligible units a known selection mechanism; stratified sampling ensures coverage of important subgroups; systematic sampling is convenient but vulnerable to periodic patterns. Convenience samples may be biased even when large. Distinguish the sampling unit from repeated measurements within that unit, because correlated repeats do not provide the same information as independent units. The population, time period and exclusions determine where conclusions can reasonably apply.
Worked example: To estimate weekly order errors, a team samples randomly within each day's orders rather than taking only Monday morning work. If it deliberately oversamples a small high-risk category, it uses appropriate population weights when estimating the overall weekly error proportion.
Common mistake: Increasing a biased convenience sample's size can improve precision around the wrong population value without correcting its selection bias.
Official syllabus reference: [1]
34. The central limit theorem concerns sample means
The central limit theorem explains why the sampling distribution of a mean can become approximately normal as sample size grows under suitable conditions. It does not make the original observations normally distributed. For independent observations with finite variance, the mean's standard error is σ/√n, estimated using sample information when necessary. Strong skew, extreme tails and dependence can demand greater care; there is no universal sample-size guarantee. Distinguish population spread from uncertainty in an estimated mean, particularly when deciding whether a method's assumptions are reasonable.
Worked example: Individual handling times are right-skewed with standard deviation twelve minutes. For independent samples of sixty-four requests, the mean has estimated standard error about 1.5 minutes. The average becomes more precise, while individual customers can still experience widely varying times.
Common mistake: Invoking the central limit theorem does not justify treating individual observations as normal when predicting individual tail probabilities or capability.
Official syllabus reference: [1]
35. Alpha, beta, power and meaningful effect size
A hypothesis test compares evidence with a stated null model. Alpha controls the planned probability of rejecting a true null under the procedure's assumptions. Beta is the probability of missing a specified alternative effect, and power equals one minus beta. Power depends on effect size, variation, sample size and test design. A p-value measures evidence under the null, not the probability that the null is true. Practical significance requires comparing the effect and its uncertainty with a meaningful business magnitude, rather than relying on statistical significance alone.
Worked example: A team plans to detect a two-minute reduction rather than any tiny difference. With estimated variation and desired alpha and power, it calculates sample size before collection. A later significant reduction of only ten seconds may still fail the business requirement.
Common mistake: Failure to reject the null does not prove equivalence; an underpowered study may simply be unable to detect the effect that matters.
Official syllabus reference: [1]
36. One-sample, independent and paired t-tests
A one-sample t-test compares a population mean with a specified value. An independent two-sample test compares means from unrelated groups; Welch's form allows unequal population variances. A paired t-test analyzes within-pair differences, such as before-after measurements on the same units. Pairing changes the unit of analysis and can remove between-unit variation. Assess independence and relevant distributional assumptions, especially with small samples or influential outliers. For paired tests, normality concerns the differences. A significant before-after change can still reflect time effects if the design lacks an appropriate comparison.
Worked example: Ten valves are tested before and after adjustment. The team calculates ten pressure differences and tests their mean against zero. Treating the twenty readings as independent would discard the valve matching and use an error structure that does not represent the design.
Common mistake: Choosing a paired test because two groups have equal sample sizes is incorrect unless observations have a meaningful matched relationship.
Official syllabus reference: [1]
37. Variance tests assess spread rather than location
A one-sample variance test can compare a normal population's variance with a specified value using a chi-squared statistic. Comparing multiple group variances requires a suitable procedure, such as an assumption-sensitive classical test or a more robust alternative. These tests address spread, not mean differences. Independence and distributional assumptions matter because classical variance tests can be sensitive to nonnormality. A nonsignificant equal-variance test does not prove variances identical. Examine graphical spread and practical consequences alongside the test, especially before selecting an ANOVA procedure.
Worked example: A filling process retains the same mean but shows a wider distribution after maintenance. A variance-focused analysis addresses whether variability increased, while a mean-only t-test could miss the operational deterioration. A time plot first checks whether the widening reflects a stable new state.
Common mistake: A nonsignificant mean comparison does not demonstrate unchanged variability, because equal centers can coexist with substantially different spreads.
Official syllabus reference: [1]
38. One-way ANOVA and planned sample size
One-way ANOVA tests whether several population means are all equal by comparing between-group variation with within-group variation. A significant result indicates at least one difference, not that every pair differs. Classical ANOVA assumes independent errors, suitable residual normality and common variance; alternatives such as Welch's ANOVA address unequal variances. Examine residuals rather than testing a pooled mixture of group observations for normality. Sample-size planning requires a meaningful difference, variability assumptions, alpha, power and design details. Follow-up comparisons should address multiplicity rather than repeatedly applying unadjusted pairwise tests.
Worked example: Three curing settings produce different average strengths. An overall ANOVA is significant, so the team uses an appropriate multiple-comparison procedure to identify supported contrasts. Before a confirmation study, it plans replication around the smallest strength difference that would justify changing the setting.
Common mistake: A significant overall ANOVA does not establish which groups differ or show that every observed ordering of sample means is real.
Official syllabus reference: [1]
39. Mann–Whitney and Kruskal–Wallis compare ranks
Mann–Whitney compares two independent groups using ranks, while Kruskal–Wallis extends rank-based comparison to several independent groups. They avoid assuming normal observations, but still require appropriate independence and measurement ordering. These tests detect distributional differences; interpreting them specifically as median or location tests requires additional shape assumptions. Different spreads or shapes can produce significance even without a simple median shift. Kruskal–Wallis is an overall test, so identifying particular differences requires suitable follow-up. Neither procedure preserves matched or blocked relationships unless the design is addressed differently.
Worked example: Two suppliers have skewed delivery times. A Mann–Whitney result is significant, but one supplier also has a much wider distribution. The team examines shapes and probabilities rather than stating that the test has proven a particular median difference without further justification.
Common mistake: Calling every rank-test result a median comparison overlooks the additional assumptions needed to interpret distributional differences as a location shift.
Official syllabus reference: [1]
40. Mood's median test trades information for robustness
Mood's median test compares independent groups by classifying observations relative to a pooled median and analyzing the resulting counts. It focuses on median-related differences while using less information than procedures that retain full ranks or measured distances. This can make it less powerful when stronger methods' assumptions are reasonable, but less sensitive to some extreme observations and shape differences. Ties at the pooled median need consistent handling. The test still requires a suitable independent sampling design, and a nonsignificant result does not establish equal group distributions.
Worked example: Service-time groups contain severe outliers and differently shaped tails. The team considers Mood's median test because its question concerns central position rather than the full distributions, while acknowledging that converting times into above-or-below counts sacrifices information and may require more observations.
Common mistake: Choosing Mood's median test does not remove the need for independent observations or justify interpreting a nonsignificant result as proven equality.
Official syllabus reference: [1]
41. Friedman preserves matched comparisons across treatments
The Friedman test provides a rank-based comparison of several treatments in a complete blocked design. Each block receives every treatment, and responses are ranked within blocks before treatment rank totals are compared. Blocking removes irrelevant between-block level differences from the comparison. Blocks should be independent, and treatment assignment or order should avoid systematic confounding where feasible. A significant result indicates an overall treatment difference, with appropriate follow-up needed to locate it. Kruskal–Wallis is not an interchangeable substitute because it assumes independent groups rather than repeated observations within blocks.
Worked example: Six printers each test three paper types in randomized order. Since every printer contributes a result for each paper, printer is a block. Friedman compares paper rankings within printers, avoiding a comparison distorted by one printer consistently producing higher jam counts.
Common mistake: Pooling all treatment observations into independent groups discards the block structure and can mistake machine-level differences for treatment evidence.
Official syllabus reference: [1]
42. Sign and Wilcoxon signed-rank tests use differences differently
A sign test uses only whether an observation or paired difference lies above or below a hypothesized value. It supports a median-oriented question with relatively few distributional assumptions, but discards magnitude information. The Wilcoxon signed-rank test ranks absolute nonzero differences and restores their signs, using more information when its symmetry assumption for the difference distribution is reasonable. Both can address paired data by analyzing within-pair differences. Zero differences and ties require the chosen implementation's appropriate handling. These are distinct from independent-group rank tests.
Worked example: A before-after study has asymmetric differences with several very large improvements. The team may use a sign test for a median-change question rather than assuming signed-rank symmetry. If differences are reasonably symmetric, signed-rank can use their relative magnitudes to improve sensitivity.
Common mistake: Choosing Wilcoxon signed-rank solely because raw before measurements are nonnormal ignores that its relevant symmetry assumption concerns the paired differences.
Official syllabus reference: [1]
43. One- and two-sample proportion inference
A one-sample proportion procedure compares a binary-outcome probability with a specified value. A two-sample procedure compares probabilities in independent populations. The numerator counts qualifying units and the denominator counts eligible units; multiple defects on one unit require another framework. Normal approximations depend on adequate expected counts, while exact or other suitable methods can address sparse data. Confidence intervals quantify effect uncertainty, and practical interpretation should include absolute percentage-point change. Paired binary outcomes require methods that recognize matching rather than treating before and after counts as independent samples.
Worked example: One line has twenty defective units among four hundred, and another has twelve among three hundred. Their estimated defective proportions are 5% and 4%, a one-percentage-point difference. A suitable comparison evaluates uncertainty rather than concluding that the smaller sample percentage proves improvement.
Common mistake: A fall from 5% to 4% is one percentage point and a 20% relative reduction; those descriptions are not numerically interchangeable.
Official syllabus reference: [1]
44. Chi-squared tables test categorical association
A chi-squared contingency-table test compares observed category counts with expected counts under independence or equal distributions. Each expected count is the row total multiplied by the column total, divided by the grand total. Observations must be appropriately independent and categories mutually exclusive. Sparse expected counts can require a different approach rather than blindly applying the approximation. A significant overall result shows association, not causation or significance in every cell. Examine standardized residuals and suitable follow-up comparisons, while respecting sampling boundaries and controlling unnecessary multiple testing.
Worked example: A table cross-classifies shift and defect type. A cell with row total eighty, column total fifty and grand total two hundred has expected count twenty under independence. Large departures contribute to the statistic, but the team investigates process differences before claiming shift causes defects.
Common mistake: Using percentages without their underlying counts loses the sample-size information needed to assess whether observed categorical differences exceed sampling variation.
Official syllabus reference: [1]
Improve Phase – Regression and Solution Development
45. Correlation measures linear association, not causation
Pearson correlation summarizes the direction and strength of a linear relationship between two numerical variables. It is unitless and ranges from minus one to plus one, but a value near zero can conceal a strong curved relationship. Outliers, restricted ranges and mixed subpopulations can substantially change it. Correlation is symmetric and does not establish which variable drives the other. A scatterplot and process knowledge should precede interpretation. Confounding variables or common time trends can generate impressive correlations without a direct causal mechanism suitable for intervention.
Worked example: Energy use and defect count rise together across days, but production volume also rises. A strong raw correlation may reflect more units being made rather than energy causing defects. Examining rates, volume and the physical process produces a more defensible improvement hypothesis.
Common mistake: Squaring a correlation may describe linear fit strength in simple regression, but it does not convert observational association into a causal effect.
Official syllabus reference: [1]
46. Simple regression, slope and residuals
Simple linear regression models the conditional mean response as β0 + β1x. The slope describes the modeled response change per one-unit predictor increase, while the intercept describes the fitted mean at zero and may lack practical meaning outside the data range. A residual is observed minus predicted response. Least squares chooses coefficients that minimize squared residuals. Predictions within the observed range are interpolation; extrapolation adds untested model assumptions. A fitted slope is not automatically causal, and residual diagnostics are needed before relying on standard errors, tests or intervals.
Worked example: A model gives predicted time = 25 − 0.5 × temperature for temperatures from twenty to forty degrees. At thirty degrees, prediction is ten minutes. An observed twelve-minute result has residual plus two minutes; predicting at eighty degrees would be unsupported extrapolation.
Common mistake: Interpreting the intercept as a real operating condition when zero lies far outside the fitted range gives the equation unsupported physical meaning.
Official syllabus reference: [1]
47. Multiple regression, curvature and collinearity
Multiple regression models a response using several predictors, so each coefficient describes a conditional association with other included terms held fixed. Correlated predictors can make individual coefficients unstable and inflate their uncertainty even when overall prediction appears strong. Curvature can be represented through terms such as x²; a model can be linear in coefficients while curved in x. Other relationships require genuinely nonlinear parameter estimation. Choose terms using process reasoning and diagnostic evidence, not only automated significance rules. Adjusted associations still require causal justification before being treated as intervention effects.
Worked example: Drying time depends on temperature and airflow, which usually increase together in production. A multiple regression attempts to distinguish their conditional relationships, but strong collinearity produces wide coefficient intervals. A well-fitting prediction does not establish a precise separate effect for each setting.
Common mistake: A high overall R-squared does not guarantee stable individual coefficients, reliable extrapolation or a causal interpretation for each included predictor.
Official syllabus reference: [1]
48. Residual diagnostics and Box–Cox transformation
Residual plots assess whether a regression leaves systematic structure unexplained. Curvature suggests a missing functional form, a funnel suggests changing variance and a time pattern suggests dependence or process change. A Box–Cox transformation uses a power family for positive responses, with the logarithm as the limiting case at zero power. It can improve variance behavior and distributional fit, but does not repair omitted mechanisms or dependence automatically. Interpret transformed-scale coefficients carefully, and recognize that inverse-transforming a fitted transformed mean does not generally recover the original-scale arithmetic mean.
Worked example: Residual spread grows with predicted cost, so a team evaluates a log-response model. The funnel diminishes, but residuals still cluster by day. The transformation addresses spread while the day pattern requires a separate explanation; simply accepting the transformed model would leave dependence unresolved.
Common mistake: A transformation that improves residual normality does not automatically correct autocorrelation, omitted variables or the causal limitations of observational data.
Official syllabus reference: [1]
49. Confidence and prediction intervals answer different questions
A confidence interval for a regression mean quantifies uncertainty about the conditional average response at specified predictors. A prediction interval for a new observation also includes individual outcome variation and is therefore wider under the same model and confidence level. Both depend on the model, sampling assumptions and predictor location; uncertainty generally increases away from well-supported regions. A two-sided interval endpoint is not automatically a one-sided percentile or guarantee. Use the interval matching the decision, and take care with back-transformation when the response has been modeled on another scale.
Worked example: At a specified order size, the mean delivery-time interval is forty-eight to fifty-two hours, while the new-order prediction interval is thirty-eight to sixty-two. Staffing estimates may use the conditional mean; an individual customer promise requires attention to the predictive spread.
Common mistake: Using a narrow mean confidence interval as the expected range for individual orders understates the variation customers will actually experience.
Official syllabus reference: [1]
Control Phase – Sustainment, SPC and Control Plans
50. 5S controls make abnormalities visible
The syllabus's 5S sequence is Sort, Straighten, Shine, Standardize and Self-Discipline. Sort removes unnecessary items; Straighten establishes clear locations; Shine combines cleaning with inspection; Standardize defines repeatable arrangements; Self-Discipline sustains use. Visual locations, replenishment boundaries and ownership help make missing tools or abnormal conditions immediately visible. An audit should assess whether the arrangement supports the work rather than merely rewarding attractive appearances. 5S controls need maintenance after layout or product changes, otherwise a once-useful standard can become an obstacle to current operations.
Worked example: A workstation assigns labeled outlines to required gauges and removes obsolete fixtures. Cleaning includes checking hoses for leaks, while a brief shift check confirms readiness. A missing gauge becomes visible before production starts rather than after an operator searches several storage cabinets.
Common mistake: Treating 5S as a one-time cleanup omits the standardization and sustained ownership that prevent the original disorder from returning.
Official syllabus reference: [1]
51. Kanban limits work through replenishment signals
Kanban authorizes production or movement in response to consumption, supporting pull rather than unrestricted production. The number of containers or signals limits work in process within the defined loop. A common planning calculation uses demand during replenishment lead time, an explicitly chosen allowance and container capacity, with rounding up to whole containers. Units must be consistent. Kanban does not remove the need for reliable replenishment, quality and capacity; excessive cards can hide problems in inventory, while insufficient coverage can cause shortages when assumptions are wrong.
Worked example: Demand is forty units per hour, replenishment takes half an hour and the chosen allowance is 20%. With ten-unit containers, required coverage is twenty-four units, so the team uses three containers. It then validates actual demand and replenishment behavior before reducing coverage.
Common mistake: Adding cards whenever a shortage occurs may conceal unstable replenishment and increase inventory without correcting the reason the loop failed.
Official syllabus reference: [1]
52. Poka-yoke prevention versus detection
Poka-yoke changes a task or system to prevent an error or detect it immediately before it propagates. Prevention controls make the incorrect action impossible or difficult under intended use, while detection controls identify a mismatch and may stop progression. Keyed fittings, sequence interlocks and part-presence checks illustrate different mechanisms. Training and reminders can support performance but depend more heavily on attention. Validate the control against realistic failure modes, wear and bypass opportunities. Its presence alone does not establish effectiveness; the intended error must be demonstrably prevented or reliably detected.
Worked example: A fixture allows a component to seat only in the correct orientation, preventing reversal. A separate sensor confirms the component is present before fastening. The team deliberately challenges both controls with representative incorrect setups to verify their mechanisms rather than relying on installation records.
Common mistake: Calling a warning label mistake-proofing can overstate its protection when the incorrect action remains easy and no immediate detection occurs.
Official syllabus reference: [1]
53. SPC sampling and control-limit interpretation
SPC uses time-ordered data to distinguish common-cause variation from signals suggesting special causes. Rational subgrouping places observations together so within-subgroup variation represents the intended short-term conditions and between-subgroup changes remain visible. Control limits are estimated from process behavior; specification limits come from requirements. Establish the baseline from suitable data, define signal rules in advance and connect them to a response plan. Excessive supplementary rules increase false alarms. Recalculating limits after every signal can absorb a real process change into the baseline before it is understood.
Worked example: A line samples five consecutive units each hour so the chart can detect changes between hours. Mixing one unit from each of five shifts into every subgroup would blend shift differences into within-group spread and make the intended comparison harder to see.
Common mistake: Adjusting a stable process after every ordinary fluctuation can increase variation through tampering rather than improve the process's underlying performance.
Official syllabus reference: [1]
54. I–MR charts for individual measurements
An individuals and moving-range chart monitors continuous observations collected one at a time when meaningful rational subgroups are unavailable. The individuals chart tracks level; the moving-range chart tracks successive absolute differences. For moving ranges of two, the usual within-process standard-deviation estimate is average moving range divided by the appropriate d2 constant. Review the moving-range behavior alongside the individuals chart. Strong autocorrelation, irregular sampling or a changing measurement system can distort interpretation. Collecting single values does not by itself guarantee that an I–MR chart's assumptions are suitable.
Worked example: A laboratory produces one viscosity result per batch, so it plots each result and the absolute difference from the preceding batch. A sudden increase in moving ranges prompts investigation of alternating recipes before the team treats the pooled limits as a stable baseline.
Common mistake: Ignoring the moving-range chart can conceal a change in short-term variability even when individual values have not crossed their limits.
Official syllabus reference: [1]
55. Xbar–R charts monitor subgroup level and range
An Xbar–R chart pairs subgroup means with subgroup ranges and is commonly used for relatively small rational subgroups. The range chart monitors within-subgroup variability, while the mean chart monitors process level. Constants depend on subgroup size, so the correct values and sampling arrangement must be used. Under the standard fixed-size construction, mean limits use the grand mean plus or minus A2 times average range. Interpret variability first because unstable within-subgroup spread undermines mean-chart limits. Subgroups should represent coherent operating conditions rather than arbitrary collections of readings.
Worked example: Every hour, a cell measures four consecutive shafts and records their mean and range. A range signal follows a loose fixture even though the subgroup average remains near target. Repairing the fixture addresses the spread problem that a mean-only chart might miss.
Common mistake: A stable subgroup mean does not imply stable variability, because unusually high and low measurements can cancel within the average.
Official syllabus reference: [1]
56. Xbar–S charts use subgroup standard deviation
An Xbar–S chart combines subgroup means with subgroup standard deviations. Standard deviation uses all observations within a subgroup, whereas the range uses only the extremes, making the S approach attractive for larger subgroups. The relevant chart constants depend on subgroup size and the chosen construction. As with Xbar–R, assess the variability chart before interpreting the mean chart's limits. Rational subgrouping and suitable independence remain essential. Switching from R to S does not solve poor sampling, mixed process states or measurement-system changes; it changes the spread estimator, not the process design.
Worked example: A production team measures twenty consecutive units at each sampling time. It uses an S chart to summarize within-sample spread and an Xbar chart to monitor level. When subgroup size changes, it revisits the applicable limits rather than reusing constants for twenty units.
Common mistake: Choosing Xbar–S solely because software offers it does not compensate for subgroups that mix unrelated operating conditions or changing product populations.
Official syllabus reference: [1]
57. p and np charts count defective units
A p chart monitors the proportion of inspected units classified as defective. An np chart monitors the number of defective units and ordinarily requires a constant sample size for directly comparable counts and standard limits. Each unit contributes a binary result, regardless of how many defects it contains. For a p chart, limits depend on sample size and narrow as n increases. The binomial framework requires appropriate independence and a coherent population; overdispersion may signal clustering, mixing or other assumptions that need attention rather than merely wider unexplained limits.
Worked example: Two daily samples both contain 8% defective units, but one includes one hundred items and the other four hundred. Their p-chart limits differ because the larger sample provides more precise evidence. Plotting defective counts on one fixed np chart would obscure the unequal denominators.
Common mistake: Using an np chart with changing sample sizes can make ordinary changes in inspection volume look like changes in process quality.
Official syllabus reference: [1]
58. u charts monitor defects per exposure unit
A u chart monitors the number of defects per unit of defined exposure when inspection quantity can vary. A unit may contain several defects, distinguishing this from p and np charts. The centerline is total defects divided by total exposure, not an unweighted average of sample rates when exposures differ. Standard Poisson-based limits depend on exposure and assume a suitable event-count mechanism. Define exposure consistently, such as metres inspected or forms reviewed. Clustering, changing opportunities or different product mixtures can produce extra variation that requires investigation.
Worked example: Inspectors find twelve surface flaws in six hundred metres of material and eighteen in nine hundred metres. Both samples have 0.02 flaws per metre. A u chart preserves that exposure basis, while comparing raw counts would incorrectly suggest worse performance in the larger inspection.
Common mistake: Replacing exposure with the number of defective units changes the meaning of a u chart and loses information about multiple defects.
Official syllabus reference: [1]
59. CUSUM and EWMA accumulate small-shift evidence
CUSUM accumulates deviations from a target using a defined reference value and decision interval, while EWMA recursively weights the current observation and previous smoothed value. Both can detect sustained small shifts that isolated-point charts may miss. In EWMA, smaller lambda retains more history, and limits must match the chosen design and time point. Parameters determine sensitivity and false-alarm behavior, so they should be selected for the monitoring purpose. Accumulated evidence can signal even when the latest individual observation is unremarkable; it still requires investigation rather than automatic causal attribution.
Worked example: With lambda 0.2, previous EWMA ten and current value twelve, the update is 10.4. Several modest increases can progressively raise this statistic. A suitable EWMA limit may signal the sustained change before any individual reading crosses its separate individuals-chart limit.
Common mistake: Requiring an individuals-chart signal before acting on EWMA or CUSUM can defeat the reason for using a small-shift detection method.
Official syllabus reference: [1]
60. Control plans connect monitoring, response and benefit sustainment
A control plan identifies critical outputs and inputs, measurement methods, sampling, limits or requirements, ownership and records. Its response component specifies actions after defined signals, including containment where appropriate, notification, investigation and criteria for resuming or releasing work. The process owner must be able to operate the controls without the temporary project team. Sustained financial evaluation compares realized benefits with ongoing control costs and checks for double counting or performance drift. Statistical stability, customer conformance and financial benefit are related but separate outcomes that should each be verified.
Worked example: After reducing seal failures, a plant assigns sampling and chart review to the shift operator, escalation to engineering and release authority to the designated quality role. Finance tracks actual scrap reduction minus recurring sensor maintenance, confirming that sustained controls still produce a net benefit.
Common mistake: A chart without assigned response actions detects unusual behavior but leaves the organization uncertain about containment, responsibility and return to normal operation.
Official syllabus reference: [1]