How do Lean and Six Sigma work together?

Lean examines the movement of work and removes delays, excess handling and other activities that do not create customer value. Six Sigma examines variation and defects using measurement and statistical analysis. A process may need both: removing a queue can shorten delivery time, while reducing variation in the preceding operation can prevent the queue from forming again. Start with the customer requirement and the observed process, then select the method that addresses the measured constraint.

Does a high capability index mean the process is under control?

No. Capability compares process output with specification limits; statistical control asks whether variation is behaving consistently over time. A capability estimate from an unstable process can be misleading because its future distribution may change. Examine a time-ordered control chart first. If it shows special-cause signals, investigate them before using a capability index to describe expected performance against specifications.

What makes Black Belt analysis different from Green Belt analysis?

Both levels use DMAIC and must connect a proposed change to customer and business outcomes. Black Belt work places greater emphasis on leading deployment, evaluating evidence across teams and using more demanding analysis. Depending on the provider, that includes regression diagnostics, designed experiments, confounding, non-normal methods and robust controls. The precise topic list differs between ASQ and IASSC, so a candidate should use the pathway for the credential they intend to earn.

Lean Six Sigma Green and Black Belt Study Guide

Start with concept 1 Choose a topic

A process can deliver acceptable output today and still be unstable enough to fail tomorrow. Lean Six Sigma gives you a way to define what customers need, measure the work that produces it, test likely causes, improve the process and keep the gains. This guide surveys four separate credentials: ASQ's Six Sigma Green and Black Belts and IASSC's Lean Six Sigma Green and Black Belts. Work through the concepts assigned to your chosen pathway, then use the others to deepen related skills. The examples use small, stated data sets so you can check each calculation and its meaning. Pay particular attention to the difference between a pattern in data and evidence of a cause, and between a process that is stable and one that meets specifications.

ASQ Six Sigma Green Belt (CSSGB)

1. Choose the improvement route

Start with the business outcome, then define a process boundary that a team and its owner can influence. DMAIC suits an existing process with a measurable performance gap: define the problem, measure its baseline, analyze causes, improve the process, and control the result. Design for Six Sigma addresses a product or process that must be designed and verified against requirements before routine operation. The choice depends on whether there is an operating process to investigate, not on how ambitious the goal sounds. Record the customer outcome, business reason, process owner, and a measure that can reveal improvement. A Green Belt may analyze and solve the bounded problem while working alongside or under a Black Belt. The ASQ Green Belt body of knowledge includes both project selection and design approaches.

Worked example: An operating packing line ships 96 accurate orders in every 100, against a hypothetical goal of 99. The outcome is packing accuracy; the packing manager owns the process. Use DMAIC to locate and test causes of the gap. A proposed returns service has no operating baseline. Its service owner should define customer requirements, design the workflow, and verify performance before launch.

Watch the distinction: Do not choose a design method simply because the proposed fix is new equipment. If an existing line is underperforming, investigate its causes before committing to that fix.

Reference: [1]

2. Map value and flow

A value stream map follows an order from request to delivery, showing processing time, queues, inventory, and handoffs. A step adds customer value when it changes the output in a way the customer wants and is done correctly the first time. Waiting, avoidable transport, and rework consume time without meeting that test. Mark them separately so a long lead time is not mistaken for long processing time. Takt time is available production time divided by customer demand over the same period; it expresses the required delivery pace. Compare each station's cycle time with takt, then investigate the slowest constrained step. Faster work elsewhere can merely increase the queue before it. A map should use observed times and quantities, since a procedure may omit actual waits or repeated work.

Worked example: A shift has 420 available minutes and demand for 70 orders, so takt is 420 ÷ 70 = 6 minutes per order. Packing takes 8 minutes per order, slower than the required pace. A 30-minute queue before packing adds lead time but no customer value. Investigate packing capacity and its causes before speeding an upstream station that would feed the queue faster.

Watch the distinction: Do not add a queue's 30 minutes to a station's cycle time and call the sum takt. Takt comes from available time and demand; the queue describes flow delay.

Reference: [1]

3. Write a workable charter

A charter makes the project testable and keeps it within a useful boundary. State the current problem with a baseline, the desired outcome with a time frame, the process start and end, and who owns the process. Choose a primary measure that directly tracks the problem. Add a consequential measure to detect harm caused by an apparent improvement, such as higher cost, slower service, or displaced defects. The charter should also name the sponsor, team, and decision points. Avoid naming a solution before analysis; doing so can turn cause finding into justification for a preferred purchase. Review the charter when evidence changes the scope or baseline. For a focused set of related topics, the ASQ Green Belt study guide can be used alongside this project logic.

Worked example: Replace “fix shipping” with: “Reduce late shipments from 18 of 200 orders per week to at most 8 of 200 within three months, from order release through carrier handoff.” The baseline is 9%; the target is 4% or less. Count late orders as the primary measure and shipping cost per order as a consequential measure. The shipping manager owns the process.

Watch the distinction: A goal of “install new software” specifies an action, not a performance result. The team could complete the installation while late shipments stay at 9% or costs rise.

Reference: [1]

4. Translate the customer voice

Customer comments are evidence of needs, but words such as “fast” and “reliable” are not yet measurable requirements. Ask what event starts the clock, what promise applies, what counts as success, and which customer groups differ. A critical-to-quality characteristic turns the answer into an observable output, such as the fraction of orders delivered by the promised date. A CTQ tree can move from broad need to specific measure; quality function deployment relates customer needs to possible design or process characteristics. Kano analysis distinguishes basic expectations from features that increase satisfaction and features that delight some customers. Weighting can help allocate attention when needs compete, but weights require credible customer input and clear scoring rules. Retain separate measures for speed and reliability when one average could conceal inconsistent service.

Worked example: Customers ask for “fast and reliable” delivery. Define reliability as orders delivered by each customer's promised date divided by eligible orders. Suppose a hypothetical priority matrix scores reliability and packaging equally on improvement opportunity, but gives them customer weights of 5 and 2. Their weighted scores are 5 and 2, so reliability leads. Keep packaging defects visible as a separate measure.

Watch the distinction: Do not define “fast” as average delivery time alone. A few very late orders can matter greatly to customers while barely changing the average.

Reference: [1]

5. Set process boundaries with SIPOC

SIPOC gives a high-level view before detailed mapping or data collection. List the suppliers that provide inputs, the inputs needed, a few major process steps, the outputs, and the customers who receive them. Set explicit start and end events; otherwise different people may count different work as part of the process. Identify the process owner separately from the customer, because the owner controls the workflow while the customer judges its output. Treat exceptions and feedback deliberately. A correction received after delivery may reveal a defect in the original output, but it is not automatically an ordinary step in the bounded process. Once the boundaries are agreed, trace handoffs and select measures at the points where errors, delays, or missing inputs may arise.

Worked example: For invoicing, begin when billing receives an approved order and end when it delivers an invoice. Sales supplies the approved order; billing uses order and price data, prepares the invoice, and sends it to the customer. The customer may later request a correction, which provides feedback about invoice quality. The billing manager owns the bounded process, even though the customer receives its output.

Watch the distinction: Do not label the customer as the process owner merely because the customer reports an error. Ownership concerns authority to change and maintain the invoicing workflow.

Reference: [1]

6. Plan work and decisions

Break a project into deliverables and activities before assigning dates. A work breakdown structure shows what must be completed; a dependency network shows what must precede what. The critical path is the longest dependent route to completion, so delay on that route delays the project unless time is recovered. A shorter parallel route has slack, though its tasks still need management. Use milestones or tollgate reviews to make decisions when evidence is ready. A RACI assignment clarifies who performs work, who is accountable for its result, who gives input, and who needs an update. Aim for one accountable decision owner even when several people contribute. Recheck the schedule when a task duration or dependency changes, because a formerly noncritical path can become critical.

Worked example: Data extraction takes 3 days, followed by 4 days of analysis, giving that path 3 + 4 = 7 days. Interviews run in parallel for 5 days. If the decision waits for both paths, the earliest review is after 7 days; interviews have 2 days of slack. Name the analyst responsible for the analysis and one sponsor accountable for the resulting decision.

Watch the distinction: Do not call the longest individual task the critical path. Dependencies matter: a 3-day task followed by a 4-day task controls this seven-day schedule.

Reference: [1]

7. Prioritize failure modes

Failure mode and effects analysis asks how a product design or an operating process could fail, what the effect would be, and how existing controls would prevent or detect it. A design FMEA examines features and design choices; a process FMEA examines execution steps and their controls. For each failure mode, describe the effect, likely cause, current prevention and detection controls, and ratings for severity, occurrence, and detection using one agreed scale. Multiplying the ratings gives a risk priority number that helps screen work. The product is ordinal and can hide a highly severe mode behind low occurrence, so review the individual ratings and any critical effect as well. After an action, reassess the relevant ratings and confirm that the control actually works.

Worked example: A process FMEA gives a mislabeled parcel severity 8, occurrence 4, and detection 6, producing RPN = 8 × 4 × 6 = 192. A second failure scores 9, 2, and 5, producing 90. The first ranks higher by RPN, but the second has greater severity. Review both effects and controls before choosing actions; these hypothetical ratings are screening inputs, not measured probabilities.

Watch the distinction: Do not treat RPN 192 as twice the real-world risk of RPN 96. Rating scales are ordinal, and multiplying them does not establish an absolute risk ratio.

Reference: [1]

8. Calculate defects and yield

Define the unit, defect, and opportunity before calculating quality measures. One unit can have several defects, so defects per unit equals total defects divided by units, while the fraction of defective units needs a count of units with at least one defect. DPMO divides defects by units times defined opportunities per unit, then multiplies by one million. Opportunities must represent real, consistently inspected chances for a specified defect. First-pass yield is the fraction of units that pass a stage without rework. Rolled throughput yield multiplies stage first-pass yields when each stage's yield is measured among units entering that stage; it estimates the fraction that passes the whole sequence first time. Do not infer first-pass yield from defect count unless you also know how defects are distributed among units.

Worked example: Inspect 100 forms for three defined field errors each and find 12 field errors. DPU = 12 ÷ 100 = 0.12 defects per form. DPMO = 12 ÷ (100 × 3) × 1,000,000 = 40,000. If a separate two-stage process has first-pass yields of 0.90 and 0.80, its rolled yield is 0.90 × 0.80 = 0.72, or 72%.

Watch the distinction: Twelve defects do not necessarily mean twelve defective forms. Several errors may occur on one form, so the defective-unit rate requires a separate count.

Reference: [1]

9. Sample the process fairly

A sample should represent the process about which you intend to conclude. Random selection gives eligible units a known chance of selection; stratification deliberately samples important groups, such as shifts, then preserves or accounts for their production shares. Record the sampling period and selection rule so convenience does not silently favor easy-to-reach units. Classify the data before choosing summaries: pass/fail is nominal attribute data, an ordered rating is ordinal, and a measured length is continuous ratio data. A single overall mean or defect rate may conceal a weak subgroup, so compare stratum results before combining them. In a proportional stratified sample, weighted and unweighted totals coincide; with deliberate oversampling, weight group estimates by their population shares when estimating the plant-wide rate.

Worked example: A plant produces 80% of its units on days and 20% at night. For a proportional sample of 50, randomly choose 0.80 × 50 = 40 day units and 0.20 × 50 = 10 night units across the study period. This represents the output mix. Choosing 50 convenient day units gives no direct evidence about night performance, even if all 50 are inspected carefully.

Watch the distinction: A larger convenience sample does not repair exclusion of a subgroup. Fifty day-shift units may be measured precisely while still giving a biased picture of plant-wide quality.

Reference: [1]

10. Check the measurement system

Observed variation combines process variation with measurement error. Repeatability describes variation when the same operator measures the same item repeatedly under comparable conditions. Reproducibility describes differences among operators or other measurement conditions. Bias is the systematic difference between the average reading and a trusted reference. A gauge repeatability and reproducibility study uses repeated measurements of representative parts by multiple operators to separate these components. Check whether the parts span the operating range and whether readings are blinded and randomized enough to reduce memory or order effects. A precise gauge can still be biased, while an unbiased average can hide poor precision. Decide whether measurement uncertainty is small enough for the intended decision, such as detecting a process shift or judging a specification boundary.

Worked example: A reference part measures 10.00 mm, while repeated gauge readings average 10.08 mm. Estimated bias is 10.08 − 10.00 = +0.08 mm. If two operators also produce noticeably different readings on the same parts, study reproducibility rather than applying only a 0.08 mm correction. A correction to the average cannot remove operator-dependent spread.

Watch the distinction: Do not interpret a stable average as proof that the measurement system is sound. Operator differences can cancel in the average while making individual product decisions unreliable.

Reference: [1]

11. Separate stability from capability

First examine observations in time order for special-cause signals and drift. Capability indices summarize a process distribution against customer specification limits; they are difficult to interpret as a future promise when the process is unstable. For a stable, approximately normal process with a meaningful within-process standard deviation, Cp = (USL − LSL)/(6σwithin) measures potential fit of spread. Cpk is the smaller of (USL − mean)/(3σwithin) and (mean − LSL)/(3σwithin), so it also reflects centering. Control limits are calculated from process behavior to detect unusual change; specifications come from requirements. A process can be stable but poorly centered, or capable on a snapshot while drifting over time. Check distribution shape, measurement quality, and the basis for sigma before acting on an index.

Worked example: Suppose stable diameter data have LSL 90 mm, USL 110 mm, mean 104 mm, and within-process σ = 2 mm. Cp = (110 − 90)/(6 × 2) = 1.67. Cpk = min[(110 − 104)/6, (104 − 90)/6] = min(1.00, 2.33) = 1.00. Spread has room within the tolerance, but the mean sits nearer the upper limit.

Watch the distinction: Do not call Cp = 1.67 proof that the process is centered or in control. Cp ignores mean position, and neither index replaces a time-order stability check.

Reference: [1]

12. Investigate patterns and relationships

A multi-vari display separates variation by position, cycle, and time or operating condition. Plan observations across each suspected source so one factor is not always paired with another. A scatterplot or regression can quantify how an output changes with a candidate input, but association does not establish the input as a cause. Confounding occurs when a third condition changes with the candidate input and could explain the relationship. Look for a pattern that repeats across machines, shifts, or lots, then propose a verification test that varies the suspected cause while holding other relevant conditions steady or accounting for them. Check the size and direction of an effect, not just a correlation statistic. An apparent overall relationship can weaken or reverse within subgroups.

Worked example: Across two machines, each morning shift has a 2% defect rate and each night shift has 8%. The repeated shift gap is a stronger lead than machine identity. Temperature is higher on nights and correlates with defects, but it is confounded with shift. Compare temperatures within shifts, or run a controlled temperature change, before assigning temperature as the cause.

Watch the distinction: Do not declare temperature the root cause because its scatterplot slopes upward. If temperature changes with shift, the plot alone cannot separate the two effects.

Reference: [1]

13. Match a test to the question

Choose a hypothesis test from the response type and how observations were collected. Measurements on the same units before and after a change form paired differences; independent groups require a between-group comparison. A binary defective outcome calls for a proportion method, while a question about spread calls for a variance method. State the null claim, alternative, significance level, and assumptions before calculating. For a paired t test, inspect whether pairs are independent and whether the distribution of differences is suitable, especially in a small sample. The test statistic is estimated mean difference divided by its standard error. A small p-value concerns compatibility with the null model; it does not measure practical importance. The practice questions can help reinforce these data-structure choices.

Worked example: The same 12 parts are measured before and after calibration. Define each paired difference as after minus before. If mean difference is 0.4 mm and its standard error is 0.1 mm, t = 0.4/0.1 = 4 with 11 degrees of freedom under paired t assumptions. The positive estimate suggests readings rose; inspect the paired differences and assess whether 0.4 mm matters operationally.

Watch the distinction: Do not run an independent-sample test on these 24 readings. It discards the known pairing and can misstate uncertainty when part-to-part variation is large.

Reference: [1]

14. Read experimental effects

A designed experiment changes chosen factors deliberately and measures a response, making causal interpretation stronger than passive correlation when execution is controlled. Specify each factor and its levels, the response, the unit of experimentation, and how run order will be randomized. Replication estimates run-to-run variation; blocking can account for a nuisance condition such as day or material lot. A main effect averages a factor's influence across the other factor's levels. An interaction means the effect of one factor changes with the other, so one overall main effect may conceal a poor setting. Inspect cell results and uncertainty before selecting a combination. Also check whether the experiment's tested range matches the proposed operating range; results outside that range require new evidence.

Worked example: At low temperature, raising speed changes yield from 80% to 82%, a gain of 2 percentage points. At high temperature, it changes yield from 85% to 94%, a gain of 9 points. The difference in speed effects is 9 − 2 = 7 points, indicating an interaction in these observed settings. Replicate randomized runs before choosing high speed and high temperature.

Watch the distinction: Do not say higher speed always adds 5.5 percentage points by averaging the two gains. The observed gain depends strongly on temperature, so the combined average hides the interaction.

Reference: [1]

15. Hand control to the owner

A control plan transfers routine monitoring and action to the process owner. Select the chart from what is measured: a p chart follows the fraction of inspected units that are defective and accommodates varying sample sizes, while other charts serve measured variables or defect counts. Establish a stable baseline, calculate the chart center and limits using the chart's assumptions, and specify sampling frequency. A control-limit signal indicates an unusual process pattern that warrants cause investigation. A customer specification breach raises a separate product or service disposition question, even if the chart shows no unusual pattern. The response plan should name who checks the measurement, contains affected output, investigates the cause, records the action, and confirms recovery. Update work instructions and training so the owner can sustain the change.

Worked example: A team inspects different numbers of orders each week and records the fraction with at least one packing defect. Use a p chart, with limits that reflect each week's sample size. If a point breaches its control limit, the owner investigates a special cause. If an order fails its customer specification, contain or correct that order under the disposition process regardless of the chart signal.

Watch the distinction: Do not treat a control limit as the customer's tolerance. A process may remain statistically stable while consistently producing an unacceptable fraction of defective orders.

Reference: [1]

ASQ Six Sigma Black Belt (CSSBB)

16. Deploy projects against strategy

Black Belt deployment connects proposed projects to strategic outcomes and scarce resources. Screen candidates for customer impact, measurable benefit, feasibility, risk, dependencies, and effects on other processes. Expected savings alone can favor a project that delays more valuable work by consuming a constrained specialist. Define who sponsors the change, who owns the affected process after handoff, and how finance will verify the benefit. Use a consistent baseline and benefit definition across the portfolio so estimates can be compared. Include intangible effects explicitly when they matter, but avoid converting uncertain claims into booked savings. Sequence projects around shared resources and operational readiness, then revisit priorities when assumptions change. The ASQ Black Belt body of knowledge places strategy, leadership, and finance within this role.

Worked example: Project A promises a hypothetical $80,000 annual saving but needs the only specialist for six months. Project B promises $60,000 and can use available staff. The $20,000 difference does not settle the choice: estimate A's specialist opportunity cost, delivery risk, and strategic fit. Ask finance to confirm what each saving means and when it would appear before ranking the projects.

Watch the distinction: Do not select A solely because $80,000 exceeds $60,000. A portfolio decision must account for scarce capacity, timing, confidence in benefits, and effects on other work.

Reference: [3]

17. Value benefits over time

Net present value converts future net cash flows to today's value using an explicitly chosen discount rate. For cash flow in year t, divide by (1 + r)^t, then subtract the initial investment. A positive NPV means the estimated discounted inflows exceed the cost under those assumptions; it is sensitive to timing, rate, and whether projected benefits become real cash effects. Separate a verified reduction in current expense from cost avoidance, which prevents a possible future expense but may never appear as a smaller existing budget line. State whether estimates are gross or net of implementation and ongoing costs. Finance should agree on baseline, accounting treatment, and verification period. Compare NPV with other project risks and constraints rather than treating one computed figure as an automatic decision.

Worked example: A hypothetical project costs $100,000 now and returns $60,000 at each of the next two year-ends. At 10% annually, NPV = −$100,000 + $60,000/1.10 + $60,000/1.10² = −$100,000 + $54,545.45 + $49,586.78 = about $4,132.23. It is positive under these cash-flow assumptions. Any avoided future expense should be described separately from verified booked savings.

Watch the distinction: Do not add two $60,000 returns and report a $20,000 NPV. That is undiscounted net cash flow; NPV accounts for when each return arrives.

Reference: [3]

18. Manage adoption barriers

A technically sound change can fail if the people who control its handoffs, schedules, or inputs cannot adopt it. Map stakeholders by their influence, exposure to the change, and current support. Then gather readiness evidence: staffing, skills, system access, supplier capacity, and likely disruptions. Match action to the barrier. Explain the reason and expected effect to those who need understanding, provide practice for a skill gap, and arrange a contingency for a capacity gap. A pilot can reveal whether the new method works across shifts and interfaces before full rollout. Assign an owner for feedback and track both adoption and process outcomes. Resistance may identify a real operating constraint, so investigate it as evidence rather than labeling it a communication problem.

Worked example: A new order handoff affects operators, supervisors, and an external supplier. Supervisors set rosters but have low support because overlap coverage is uncertain. Meet them with the proposed schedule, test a pilot shift, and define backup coverage before rollout. Train operators on the new handoff and confirm the supplier can receive it. Track missed handoffs as well as attendance at training.

Watch the distinction: Do not count a general announcement as evidence of readiness. People can understand the change and still lack the staffing or authority needed to perform it.

Reference: [3]

19. Lead a working team

Choose leadership behavior according to the team's skill, confidence, conflict, and decision urgency. Direct when a task or safety-critical procedure needs clear instruction; coach when people need both direction and development; support competent members facing uncertainty or disagreement; delegate when ownership and capability are established. For contested causes, structure discussion around evidence. Nominal group technique lets members generate or rank ideas independently before group discussion, reducing the influence of a dominant voice. Consensus can build commitment, while multivoting can narrow options, but neither method proves a cause. Agree on a verification test, its data, and a decision rule. Remote teams benefit from written prework and a clear meeting record so absent or quiet members can contribute and the chosen action remains traceable.

Worked example: A skilled remote team disputes whether defects come from material or setup. Use a supporting approach: ask members to rank the evidence independently, compare their rankings, and identify what evidence is missing. They agree to test both materials under a controlled setup, then both setups with one material. The resulting defect data, rather than a vote, will decide which cause is supported.

Watch the distinction: Do not use majority voting to declare a root cause. Voting records preference, while a cause requires evidence that distinguishes it from competing explanations.

Reference: [3]

20. Build a line of sight to outcomes

A line of sight connects a customer need to a process measure, an action people can take, and a business result. A critical-to-quality measure expresses the customer outcome, such as correctly resolved complaints. A leading behavior indicator records whether a practice expected to influence that outcome is performed. A lagging indicator records the outcome after enough work has passed through the process. Keep their denominators and definitions stable so changes are interpretable. A rise in the leading indicator is evidence of adoption, not proof of benefit; check whether the lagging result changes and whether volume or case mix changed. If checks rise but complaints do not fall, examine check quality, the assumed causal link, and other drivers before intensifying the same activity. A balanced scorecard can show related customer, process, and financial effects together.

Worked example: Suppose a team checks first-contact information on 800 of 1,000 orders, then on 950 of 1,000 after training: the leading rate rises from 80% to 95%. Verified complaints fall from 30 to 20 per 1,000 comparable orders, a decline of 10 per 1,000. That supports continued monitoring, while a review should still test whether the checks addressed the complaints.

Watch the distinction: Do not call a higher check-completion rate the final customer result. It measures behavior; the complaint rate shows whether the desired outcome moved.

Reference: [3]

21. Find the true flow constraint

Work in process (WIP) includes units being worked on and units waiting. In a stable flow with consistent units, Little's Law gives average lead time = average WIP ÷ throughput. Compare that lead time with touch time and map the queues between steps. A large gap points to waiting, handoffs or rework. Observe where work accumulates, then compare each step's effective output rate with demand. A machine's overall equipment effectiveness combines availability, performance and quality; a busy machine can still release too few good units. Rework creates a hidden flow of repeat visits, so count those visits when locating the constraint. Increasing speed at a step that is not constraining throughput may only enlarge the next queue. The ASQ Black Belt study guide places this flow diagnosis alongside the other Measure skills.

Worked example: A stable line completes 20 good units per hour and averages 60 units in process. Little's Law gives 60 ÷ 20 = 3 hours of lead time per unit. If direct work takes 10 minutes, about 170 minutes of average lead time occurs elsewhere. A queue before inspection and units returning for repair deserve investigation before the team speeds the 10-minute step.

Watch the distinction: High equipment utilization does not prove a step is the constraint. Check its good-unit output rate, the queues around it and any rework loop before investing in speed.

Reference: [3]

22. Protect data integrity

A defensible data plan specifies the characteristic, unit, operational definition, collection point, time window and sampling method before observations begin. Random selection reduces selection bias; stratification preserves representation when shifts, products or suppliers differ. Record missing and invalid values rather than silently replacing them. A calibrated gauge needs a reference with traceable identity, an appropriate range and recorded conditions. Compare repeated readings with reference values to detect bias, and check whether the bias changes across the operating range. Resolution must be fine enough to distinguish differences that matter to the decision. After repair or adjustment, document what changed and determine which earlier measurements may be affected. More samples narrow sampling uncertainty, but they cannot remove systematic measurement bias.

Worked example: After repair, a gauge averages 50.12 mm on a traceable 50.00 mm reference. Estimated bias is 50.12 − 50.00 = +0.12 mm. Record the reference, date and conditions, then check other values across the working range. Review production measurements taken since the last satisfactory check; their affected interval depends on the gauge's history. Measuring additional production parts with this gauge would retain the bias.

Watch the distinction: A large, precisely collected sample can still be wrong when the gauge has systematic error. Fix and characterize the measurement system before using its readings to judge process capability.

Reference: [3]

23. Handle non-normal capability

Capability asks how a process distribution relates to customer specifications, so examine time order before fitting any distribution. A special cause or drift makes a single long-run capability estimate hard to interpret. Next inspect a histogram, probability plot and process knowledge for skew, multiple operating modes and outliers. Normal-model Cp and Cpk summarize distances in standard-deviation units; their usual tail interpretation depends on a suitable normal model. For stable non-normal data, a justified transformation such as Box-Cox or a fitted distribution can support percentile-based comparisons with specification limits. Validate the fit, especially near the relevant tail. Separate shifts or product families when they represent distinct processes. The ASQ Black Belt body of knowledge includes non-normal capability; the model choice still rests on the observed process.

Worked example: Five delivery times are 1, 2, 2, 3 and 12 days. Their mean is 20 ÷ 5 = 4 days, while the median is 2 days. The 12-day case makes the sample right-skewed and may reflect a separate delay mechanism. Investigate it and collect enough time-ordered data before fitting a tail model or estimating performance against a delivery specification.

Watch the distinction: A transformed dataset is not automatically capable. First establish stability, then verify that the transformation or fitted distribution represents the tail that reaches the specification limit.

Reference: [3]

24. Choose a probability model

Choose a probability model from what each observation counts. A binomial variable counts defective units in a fixed number of trials when each unit has a defined pass/fail outcome and trials are suitably independent with a common defect probability. A Poisson variable counts events over a fixed area, time or opportunity exposure when events occur independently at a roughly constant rate. One item can contribute several Poisson events but only one binomial defective-unit outcome. Lifetime is continuous: an exponential model assumes a constant failure rate, while a Weibull model can represent rates that change with age. Check exposure, independence, clustering and changing conditions before applying any model. A model supports estimates and uncertainty calculations; it cannot rescue inconsistent definitions of a defect or an inspection opportunity.

Worked example: Suppose 100 independently inspected units each have a hypothetical 0.05 probability of being defective. A binomial model gives an expected defective count of 100 × 0.05 = 5 units. If plates instead average two scratches per plate under a suitable constant-rate model, 100 plates have an expected 200 scratches. Several scratches can occur on one plate, so the counts answer different questions.

Watch the distinction: Do not use a binomial count for scratches merely because plates are inspected individually. Binomial counts plates classified defective; an event-count model can count multiple scratches on one plate.

Reference: [3]

25. Interpret uncertainty and power

A confidence interval describes uncertainty in an estimated population quantity, such as a mean improvement. A prediction interval concerns one future observation and is usually wider because it includes individual variation. A significance test asks whether the data conflict enough with a stated null model; it does not say whether the effect is large enough to matter operationally. Set the practical target before reviewing results. A type I error is rejecting a true null hypothesis, controlled in part by the chosen alpha. A type II error misses a specified real effect; power is one minus that probability. Power depends on effect size, variation, sample size and the test design. When a result is inconclusive, inspect the interval and study power before declaring that two processes are equivalent.

Worked example: A trial estimates a mean time saving of 3 minutes, with a hypothetical 95% confidence interval from 1 to 5 minutes. The interval excludes zero, supporting a positive mean saving under the study assumptions. Yet a required saving of at least 4 minutes is not established because values below 4 remain plausible. One future job's saving would need a separate, generally wider prediction interval.

Watch the distinction: A statistically significant improvement need not meet the business target. The interval gives the plausible effect range; compare that range with the target rather than relying on a p-value alone.

Reference: [3]

26. Compare several groups

One-way analysis of variance compares the means of three or more groups through a ratio of between-group variation to within-group variation. Its overall null hypothesis says all group population means are equal. A significant F test indicates evidence that at least one differs, but it does not identify a pair. Examine independent sampling, residual shape and spread across groups; unequal variances or skew may call for another method or a justified transformation. Plot the observations as well as the means, since a supplier with a favorable average may have unacceptable dispersion or outliers. If the overall test warrants follow-up, use planned contrasts or a multiple-comparison procedure that accounts for repeated pairwise testing. Choose a supplier using the relevant specification, cost and reliability measures, not the mean alone.

Worked example: Three suppliers have sample means of 10, 11 and 15 strength units. Assume a valid one-way ANOVA gives p = 0.01 with a prespecified alpha of 0.05. Since 0.01 < 0.05, reject equal population means. This result identifies no particular pair. Inspect residuals and use an appropriate follow-up comparison before claiming supplier C differs from A or B.

Watch the distinction: A significant ANOVA result does not prove that every supplier differs from every other supplier. Pairwise claims require follow-up comparisons and attention to multiple testing.

Reference: [3]

27. Validate a regression model

A regression coefficient describes an expected change in the response per unit change in a predictor, holding other model terms fixed. Interpret it only within the range and operating conditions used to fit the model. Check residuals against fitted values, predictors and time: curvature suggests a missing nonlinear term, widening spread suggests nonconstant variance, and time patterns suggest dependence. Check unusual influential observations before allowing them to determine the slope. A fitted relationship may predict well without showing that changing the predictor will cause the response to change. For a decision, compare predicted outcomes and their uncertainty at feasible settings, then verify the intervention. A confidence interval addresses the mean response at a setting; a prediction interval for one future case also includes case-to-case variation.

Worked example: A fitted equation is delay in hours = 2 + 0.5 × queue length in jobs. At six jobs, predicted delay is 2 + 0.5 × 6 = 5 hours. If high-queue residuals are consistently positive, actual delays there exceed the straight-line predictions. The team should revise or restrict the model before using its six-job estimate for staffing decisions.

Watch the distinction: A high correlation or small slope p-value does not validate every prediction. Residual patterns and the observed predictor range can reveal systematic error where decisions matter most.

Reference: [3]

28. Design a fair experiment

An experiment estimates a setting's effect by deliberately changing a factor while protecting the comparison from nuisance variation. Randomize run order to reduce confounding with time trends. Block on a known nuisance source, such as production day, so each setting is compared under similar conditions. Replication means new independent experimental runs; repeated readings of one run estimate measurement variation but do not replace replication. Define the response, factor levels, measurement method and intended effect size before running the study. With multiple factors, plan for interactions: the effect of one setting may depend on another. Review residuals and practical constraints before selecting an operating setting. An experiment with one observation per setting per block can estimate day-adjusted differences, though its precision remains limited by the available runs.

Worked example: A team tests settings A and B on each of four production days, randomly choosing which runs first each day. That produces eight runs and four within-day A-versus-B contrasts. If the daily differences, B minus A, are 2, 3, 1 and 2 yield points, their mean is (2 + 3 + 1 + 2) ÷ 4 = 2 points. Day blocking protects that comparison from day-level material changes.

Watch the distinction: Running all A trials before all B trials confounds setting with time. If material or temperature changes during the study, the apparent setting effect may be a time effect.

Reference: [3]

29. Select advanced controls

Choose a chart from the response type and sampling structure. Consecutive numeric observations can use an individuals and moving-range chart; small rational subgroups of numeric readings can use an X-bar and R chart. Fraction defective uses p or np charts, while defect-event counts use c or u charts according to exposure. Rational subgroups capture short-period variation within groups while making shifts between groups visible. For short production runs, a short-run chart can standardize unlike products against their own targets and scales. Control limits describe expected process behavior estimated from process data; specifications state external requirements. Investigate a special-cause signal before recalculating limits. When improvement reduces process variation, repeat measurement-system analysis because unchanged gauge error becomes a larger share of the observed spread.

Worked example: Five consecutive diameter readings are taken each hour, supporting an X-bar and R chart if within-hour conditions are reasonably alike. Suppose process standard deviation falls from 4 to 2 units while gauge standard deviation stays at 1 unit. Under independent additive errors, gauge variance rises from 1 ÷ (16 + 1) = 5.9% to 1 ÷ (4 + 1) = 20% of observed variance. Reassess the gauge.

Watch the distinction: A chart that uses subgroup means does not replace scrutiny of within-subgroup ranges. A stable mean chart can coexist with changing short-term spread or inadequate measurement precision.

Reference: [3]

30. Design for constraints and robustness

Design-stage work starts with customer needs and translates them into measurable targets before the process exists. Use design validation to check that a proposed product or service meets those needs under realistic conditions; operating-process DMAIC instead diagnoses and improves an existing process. Robust design seeks performance that changes little when noise factors, such as temperature or material variation, change. Tolerance design sets acceptable ranges for components while considering assembly behavior, inspection capability and cost. Statistical tolerancing can estimate the spread of an assembled characteristic when component variations are suitably modeled; a worst-case check may still matter for critical failures. Compare alternatives across cost, manufacturability, testability and maintenance, then verify the chosen design at realistic extremes. Nominal performance alone is insufficient when customers experience varying conditions.

Worked example: Suppose an output must stay between 95 and 105 units over a stated temperature range. At nominal temperature, designs A and B both deliver 100. Across the range, A delivers 93 to 107 and B delivers 98 to 102. B remains within the hypothetical specification while A does not. If B costs $2 more per unit and can be produced and tested reliably, compare that cost with the avoided failures before selection.

Watch the distinction: Passing a nominal-condition test does not establish robustness. Check the range of expected use and production variation, then confirm that the proposed tolerances can be built and measured.

Reference: [3]

IASSC Lean Six Sigma Green Belt (ICGB)

31. Trace the output to inputs

Write the outcome of interest as Y and plausible process inputs as x variables in Y = f(x). This notation is a search strategy, not a claim that every listed input is causal. Define Y in customer terms and measure its baseline, then map the process to locate candidate drivers. Include the voice of the customer, business and employee: a faster promise is useful only if cost and work conditions can support it. Prioritize inputs using process knowledge and data, then test the strongest candidates while holding competing explanations in view. Distinguish a controllable input, such as batch size, from an external condition that may need monitoring. The IASSC Green Belt syllabus places this problem-solving view within Define.

Worked example: A depot defines Y as the weekly percentage of orders delivered after the promised date. In a hypothetical week, 18 of 200 orders are late, so Y = 18 ÷ 200 = 9%. Batch size and pickup time are candidate x variables. Compare lateness by batch and pickup window while tracking transport cost and shift coverage; an association alone does not establish which change will help.

Watch the distinction: Listing inputs in a process map does not prove they drive the output. Measure variation in both Y and candidate x variables, then test alternatives before changing schedules.

Reference: [5]

32. Recognize waste at the source

The seven waste categories listed in the IASSC Green Belt syllabus are overproduction, correction, inventory, motion, overprocessing, conveyance and waiting. Classify an observed activity by what happens, then ask why it occurs; the label does not supply the cause. Producing ahead of demand often creates inventory and handling. Correction consumes capacity because work must be repeated. Motion concerns people's movement, while conveyance concerns moving material. Overprocessing adds work the customer does not require. A 5S workplace routine sorts unneeded items, arranges what remains, cleans while inspecting, standardizes the arrangement and sustains it through regular practice. It helps make missing tools and abnormal conditions visible. Measure flow and defects afterward to see whether the change improved the outcome; a tidy area can still contain a faulty process.

Worked example: A line makes 120 kits for 100 confirmed orders, leaving 120 − 100 = 20 excess kits. That is overproduction and may create inventory. Removing obsolete parts and marking component locations can reduce searching motion. If five of the 100 required kits contain the wrong component, those defects need a separate cause investigation; clearer storage may help, but the count alone does not prove it.

Watch the distinction: A clean, labeled work area is not proof that waste has been eliminated. Track the specific excess output, search time, waiting or defects the 5S change was meant to reduce.

Reference: [5]

33. Define benefits that can be captured

A credible business case starts with a stable baseline, a defined defect or rework unit, its cost and the expected volume. Cost of poor quality can include scrap, repair, returns and other costs caused by failing to meet requirements, but avoid counting the same loss twice. State whether the proposed benefit is a booked cost reduction, released capacity or avoidance of future spending. Multiply the change in event rate by comparable volume and a supportable cost per event; then subtract the cost of implementing and sustaining the change. Agree with finance how benefit will be observed, when it will be recognized and who will verify it. A lower count can result from lower production, so compare rates when volumes change. Use a range when either future volume or unit cost is uncertain.

Worked example: A process averages 40 reworks per month at a hypothetical $25 each, costing 40 × $25 = $1,000 monthly. A target of 15 reworks at comparable volume implies 25 fewer reworks and 25 × $25 = $625 monthly gross avoidance. If the control costs $100 per month, expected net benefit is $525 monthly, subject to verifying both the rework reduction and actual cost treatment.

Watch the distinction: Do not call a target a realized saving. Capture the post-change event rate, confirm comparable volume and cost, and distinguish avoided work from expenses actually removed.

Reference: [5]

34. Prioritize possible causes

A fishbone diagram gathers possible causes by process category, making gaps in the team's thinking visible. It does not rank or validate those causes. An X-Y matrix relates candidate inputs to important outputs: weight each output by customer importance, score each input's plausible relationship and sum the weighted scores. Use the ranking to decide what to measure first. Failure mode and effects analysis then asks how a process can fail, what the effects are, what controls exist and where prevention or detection should improve. Severity deserves separate attention because multiplying ordinal ratings into a risk priority number can conceal a serious failure with a modest product. Revisit the rankings after data collection. A high matrix score is a hypothesis for testing, while a severe failure may demand controls even if it is not the leading driver of average performance.

Worked example: Suppose seal strength has an importance weight of 5. The team scores temperature's relationship at 9 and label placement's at 3, giving weighted scores of 5 × 9 = 45 and 5 × 3 = 15. Measure temperature first for seal strength. If misplaced labels create a severe identification failure, retain that mode in the FMEA and review its prevention separately.

Watch the distinction: An X-Y score is not measured evidence of causation. Nor does a low risk priority number automatically clear a high-severity failure; inspect severity and controls explicitly.

Reference: [5]

35. Read distribution shape

Start with a time plot to see whether observations come from one stable process, then use a histogram or probability plot to assess shape. The mean uses every value and is sensitive to an extreme observation; the median marks the middle and is more resistant. Standard deviation measures spread around the mean, while an interquartile range describes the middle half without letting the tails dominate it. A long upper tail may reflect occasional delays, a separate operating mode or data error. Check the underlying records before deleting an unusual value. Normality diagnostics are aids, not substitutes for a process explanation: with small samples they have limited power, and with large samples they may flag minor departures. Select summaries and any later model according to the decision, especially whether tail performance matters.

Worked example: Cycle times in minutes are 4, 5, 5, 6 and 20. The sum is 40 minutes, so the mean is 40 ÷ 5 = 8 minutes; the ordered middle value gives a median of 5 minutes. The 20-minute job pulls the mean above the typical observation. Check whether it involved rework or a recording error before summarizing normal operation.

Watch the distinction: Calling the mean 'typical' can mislead when a long tail drives it. Report shape and a suitable center measure, then investigate exceptional observations before excluding them.

Reference: [5]

36. Detect bias across the range

Bias is the average reading minus a reference value at a stated point. Linearity describes whether that bias changes across the measurement range, while stability describes whether performance changes over time. Check several traceable reference values spanning the range used in production, repeat measurements at each value and plot bias against reference size. A consistent offset suggests one kind of correction; a changing offset needs a range-sensitive calibration or repair. Repeat checks after time or use to detect drift. Also inspect repeatability, because noisy readings make a bias estimate uncertain. The reference, operator, environment and date should be recorded so later decisions can identify affected measurements. A two-point comparison can flag a problem, but more points are needed to establish its shape across the full range.

Worked example: A gauge reads 10.02 mm on a 10.00 mm reference and 100.20 mm on a 100.00 mm reference. Biases are +0.02 and +0.20 mm, a rise of 0.18 mm over 90 mm of reference span. The two-point bias slope is 0.18 ÷ 90 = 0.002 mm per mm. Check intermediate references; subtracting one constant offset would leave range-dependent error.

Watch the distinction: Good repeatability does not establish accuracy. A gauge can give nearly identical readings every time while its bias grows with part size or drifts over time.

Reference: [5]

37. Study variable and attribute agreement

Numeric measurements and pass/fail classifications need different measurement-system studies. For numeric data, repeatability is variation when the same operator measures the same item repeatedly; reproducibility concerns differences among operators or conditions. A crossed gauge R&R study deliberately repeats measurements across representative parts and operators, so the components can be separated. For attributes, have inspectors classify representative items independently, preferably including difficult cases and a defensible reference classification. Examine agreement between inspectors, each inspector's repeat agreement and agreement with the reference. Raw agreement can be inflated when almost every item belongs to one category, and it does not establish correctness. Review false accepts and false rejects separately if their consequences differ. Improve definitions, training or inspection conditions before using classifications to estimate process defect rates.

Worked example: Two inspectors agree on 18 of 20 pass/fail calls, so raw agreement is 18 ÷ 20 = 90%. They disagree on two items. A reference review shows inspector A correct on 17 of 20 and B correct on 19 of 20; their respective reference agreements are 85% and 95%. The shared calls alone hid that difference, so inspect the disputed cases and classification rule.

Watch the distinction: A 90% inter-inspector agreement figure does not mean decisions are 90% correct. Inspectors may agree on the same wrong call; compare both with a sound reference.

Reference: [5]

38. Assess attribute performance

Define the unit, defect and opportunity before calculating attribute metrics. A unit can have more than one defect, while first-pass yield counts a unit as passing only if it exits the stage without a defect or rework. Defects per unit divides defect events by units. Defects per million opportunities divides events by units times defined opportunities, then scales by one million; the opportunity definition must remain consistent across comparisons. A defective-unit proportion instead counts units with at least one defect. These measures answer different questions, especially when defects cluster on a few units. Check whether an opportunity can genuinely fail on every unit and whether inspection finds defects consistently. Use Lean Six Sigma practice questions to rehearse the denominator choice as carefully as the arithmetic.

Worked example: A team inspects 200 forms, each with four defined fields. It finds 16 field errors, distributed across 14 forms. There are 200 × 4 = 800 opportunities, so DPMO = 16 ÷ 800 × 1,000,000 = 20,000. First-pass yield is (200 − 14) ÷ 200 = 93%, assuming those 14 forms are the only ones requiring correction.

Watch the distinction: Do not divide defect events by defective forms when calculating DPMO. Use all defined opportunities; use the count of affected forms for defective-unit rate or first-pass yield.

Reference: [5]

39. Locate variation with multi-vari data

A multi-vari study samples deliberately across suspected sources of variation rather than pooling everything into one overall spread. Positional variation occurs within a unit, such as left versus right edge. Cyclical variation repeats with a machine cycle, cavity or batch position. Temporal variation appears across hours, shifts or days. Build a sampling plan that records each source on the same units where possible. Paired positions help isolate within-unit differences from unit-to-unit differences. Plot the readings in their natural order and compare the size and consistency of each contrast. A recurring contrast directs the next investigation, but it does not prove the physical cause. Check the measurement method too, since a gauge setup that changes with position could mimic a process pattern.

Worked example: On two day-shift parts, left/right readings are 8/10 and 9/11 units. On two night-shift parts, they are 8.5/10.5 and 9.5/11.5. Every right edge exceeds its paired left edge by 2 units. Overall day and night means are 9.5 and 10 units, a smaller 0.5-unit difference. Investigate position first while checking how each edge is measured.

Watch the distinction: Pooling all readings into one average hides a consistent within-part gap. Preserve part, position and shift identifiers so you can see which source dominates.

Reference: [5]

40. Use sampling distributions

A sample mean varies less than an individual observation because independent observations partly average one another out. If the population standard deviation is σ and the sample contains n independent observations, the standard error of the mean is σ/√n. Under a finite-variance population, the central limit theorem says the standardized distribution of sample means approaches normality as n grows. It does not turn the population of individual measurements normal. Strong skew, heavy tails, dependence or an unrepresentative sample can make a normal approximation poor at a given n. For sampling without replacement from a small finite population, a finite-population correction may also matter. Use the standard error for inference about a mean, while using the original process spread to describe individual outcomes and customer risk.

Worked example: Assume independent random sampling from a process with population standard deviation 12 minutes. For n = 36 jobs, the sample mean's standard error is 12/√36 = 2 minutes. Thus repeated 36-job averages vary less than individual jobs, whose standard deviation remains 12 minutes. The calculation does not show that any single future job will lie within 2 minutes of the mean.

Watch the distinction: The central limit theorem concerns the sampling distribution of the mean. It does not make individual observations normal or reduce their actual process variation as sample size grows.

Reference: [5]

41. Test normal-data questions

Start with the response and the sampling structure. A one-sample t test compares one population mean with a stated target. A two-sample t test compares independent groups; paired observations require a test on their within-pair differences instead. One-way ANOVA asks whether at least one of three or more group means differs. Check independence from how units were sampled and assigned. Inspect each group's distribution for severe skew or outliers, especially with small samples. An equal-variance two-sample t test and conventional ANOVA assume comparable group variances; Welch's versions relax that requirement. A significant ANOVA result does not identify the differing groups without follow-up comparisons. State the null, alternative, significance level and practical improvement needed before looking at results. The ICGB syllabus includes these normal-data tests.

Worked example: Suppose two independent groups of ten units have means of 52 and 48 units, and the estimated standard error of their difference is 1.5 units. Testing equal means gives t = (52 - 48)/1.5 = 2.67. If equal variances are defensible, degrees of freedom are 10 + 10 - 2 = 18, giving a two-sided p-value near 0.016. At a hypothetical 0.05 significance level, the mean difference is statistically detectable; its operational value still needs assessment.

Watch the distinction: A t statistic of 2.67 does not supply a universal p-value. The degrees of freedom, test direction and variance method affect that value, while statistical significance says nothing by itself about useful improvement.

Reference: [5]

42. Choose a rank-based comparison

Choose a rank-based test by first asking whether observations are independent or linked. For two independent groups, Mann-Whitney compares their relative ranks; for three or more independent groups, use Kruskal-Wallis. For two paired conditions, the sign test uses only the direction of each nonzero difference. Wilcoxon signed-rank also uses ranked difference magnitudes, so examine whether the paired-difference distribution is reasonably symmetric. With three or more conditions measured on the same subjects or blocks, Friedman ranks conditions within each block. These tests can be useful when a normal-data model is unconvincing, but they do not erase design problems or automatically test a difference in medians. Interpret each result in light of distribution shape, pairing and what was randomized. The IASSC Green Belt guide provides a broader route through the related analysis topics.

Worked example: Six operators each try settings A, B and C once, producing 18 measurements. Each operator supplies a block of three linked observations, so rank the settings within each operator and use Friedman to assess a setting effect. A low Friedman p-value would justify examining which settings differ, with an appropriate follow-up method. Kruskal-Wallis would treat all 18 measurements as independent and discard the operator-level pairing.

Watch the distinction: Skewed data do not make every rank test interchangeable. Mann-Whitney and Kruskal-Wallis require independent groups; applying either to repeated measurements can misstate uncertainty because the same person's results are correlated.

Reference: [5]

43. Test counts and proportions

A binary unit outcome, such as defective or acceptable, leads to a proportion question. Compare two independent suppliers with a two-proportion test when sampling and expected-count conditions support its approximation. A contingency-table chi-square test examines association between categorical variables and extends naturally beyond two groups or outcome classes. Both approaches require counts of units, not percentages alone. Check that each unit contributes to only one cell, sampling is reasonably independent and expected counts are adequate for the chosen approximation. If expected counts are sparse, an exact method may be more suitable. A small p-value indicates that the observed difference is difficult to explain under the stated null model; it does not establish why the suppliers differ. Assess the size of the rate difference and possible differences in product mix before acting.

Worked example: Supplier A has 8 defective units among 100 and B has 18 among 100: rates are 0.08 and 0.18, a 0.10 gap. Under equal rates, the pooled estimate is 26/200 = 0.13. The standard error is √[0.13 × 0.87 × (1/100 + 1/100)] ≈ 0.0476, so z ≈ -2.10 and the two-sided p-value is about 0.036. Expected defective counts are 13 per group; investigate mix before attributing the gap to supplier practice.

Watch the distinction: A ten-percentage-point sample gap is not automatically a supplier effect. Sampling uncertainty, nonindependent units and unequal product mixes can change the conclusion, even when the arithmetic for the two observed rates is correct.

Reference: [5]

44. Build and diagnose a prediction

Multiple regression estimates how an outcome changes with one predictor while holding the other included predictors fixed. Define units and code indicator variables explicitly; a rework flag coded zero or one adds its coefficient only for flagged jobs. Before using predictions, plot residuals against fitted values and each predictor. Curvature, increasing spread, runs over time or influential observations can make a seemingly strong fit unreliable. If a positive response has a nonlinear relationship or variance grows with its level, a transformation such as a log or Box-Cox transformation may help, provided predictions are interpreted on the original scale correctly. Check whether the requested predictor combination lies inside the observed data range. A confidence interval describes uncertainty in the conditional mean; a prediction interval includes variation of an individual future case and is wider.

Worked example: Suppose fitted cycle time in minutes is 4 + 0.3 × queue length + 2 × rework flag. For ten queued jobs and flag 1, the prediction is 4 + 3 + 2 = 9 minutes; without rework it is 7 minutes at the same queue length. A single future job can deviate from 9 minutes, so use a prediction interval for its time. Residual curvature would call for model revision before relying on either interval.

Watch the distinction: A high overall R-squared does not validate a particular prediction. Residual curvature or a queue length outside the observed range can make the nine-minute estimate unreliable despite apparently precise coefficients.

Reference: [5]

45. Sustain a Lean improvement

A pull control should specify what authorizes production or replenishment, the quantity released and who reacts when the signal fails. Kanban cards can limit work in process, while mistake-proofing prevents or exposes an error at its source. Neither substitutes for checking whether the process remains statistically stable. Choose a chart from the data structure: a u chart monitors defects per inspected unit when each unit may have multiple defects and inspected volume varies; a p chart monitors the fraction of units defective. Set sampling frequency and chart limits from a suitable stable baseline. A signal triggers a defined investigation, containment decision and record of cause; changing card quantities or chart limits immediately can hide deterioration. Review service, inventory and defect measures together so a local gain does not create another problem.

Worked example: A card authorizes ten replacement units when a bin empties. Weekly inspections find 12 defects across 100 units, then 15 across 150: u values are 0.12 and 0.10 defects per unit. The unequal volumes favor a u chart with limits adjusted for each sample size. If a point signals, the owner checks data and the replenishment process before changing the ten-unit card quantity.

Watch the distinction: Treating defects as defective units can select the wrong chart. One inspected unit may have several defects; a u chart tracks defect counts per unit, while a p chart tracks units classified defective.

Reference: [5]

IASSC Lean Six Sigma Black Belt (ICBB)

46. Verify strategic benefits

Benefit verification starts with a baseline that is comparable to the post-change process. Define the defect, exposure unit, time period, product mix and cost per event before reporting savings. Counts alone can fall because output fell, so normalize them as rates and calculate what the baseline rate would predict at current volume. Then examine whether material, staffing, customer mix or measurement changed alongside the project. Finance should agree how reduced expense, avoided future expense and released capacity are recorded; capacity becomes a booked saving only if costs actually leave or additional value is realized under an agreed method. Track adverse effects such as added inspection or overtime. A credible business result links the changed process, verified performance and financial records, rather than multiplying an optimistic target by an assumed unit cost.

Worked example: Scrap drops from 50 of 1,000 units to 30 of 500 units. The baseline rate is 50/1,000 = 5%; the later rate is 30/500 = 6%. At the old rate, 500 units would produce 25 scrap units, five fewer than observed. The lower count therefore reflects reduced volume and masks a worse rate. Finance should verify comparable mix and actual scrap costs before recording any gain.

Watch the distinction: Reporting twenty fewer scrap units as savings ignores the halved production volume. On a comparable per-unit basis, scrap worsened by one percentage point; the count reduction cannot establish a process benefit.

Reference: [7]

47. Expose queues in the value stream

A value stream map puts processing steps, queues, inventory and information signals on one timeline. Use it to separate touch time from elapsed lead time and to see where work waits or returns for rework. Little's Law relates average work in process, throughput and lead time for a reasonably stable system using consistent boundaries and units: WIP = throughput × lead time. A large queue may reflect batch release, uneven arrivals, constrained capacity or rework; observe the actual flow before choosing a remedy. Speeding a nonconstraint step may only build more inventory ahead of the constraint. Compare arrivals with departures, map where cases accumulate and test whether reducing batch size or stabilizing the bottleneck lowers WIP without harming service. Recalculate the whole-stream measure after the change.

Worked example: A stable process completes 40 cases per day and averages 200 cases between intake and completion. Little's Law gives average lead time = 200/40 = 5 days. If each case receives 20 minutes of touch work, queueing and other delays dominate elapsed time. Cutting touch work by one minute saves 40 minutes of labor per day at this throughput, but cannot by itself explain away roughly five days of lead time.

Watch the distinction: Little's Law does not identify the cause of a queue. The five-day figure is an average under stable-flow and consistent-boundary assumptions; inspect arrivals, bottlenecks and rework before selecting a fix.

Reference: [7]

48. Select inputs for validation

An X-Y matrix ranks candidate process inputs against several critical-to-quality outputs. Give each output an importance weight, then score the strength of each input-output relationship using a consistent scale. Multiply each relationship score by its output weight and sum across outputs. This makes assumptions visible: an input that strongly affects throughput may also worsen defects, so a single total does not capture direction or trade-offs. Record the evidence behind high scores and include process owners who understand different outputs. Use the ranking to plan measurement, regression or an experiment, not to declare a root cause. If score changes within a plausible range reverse the order, treat the ranking as fragile. A severe safety or customer requirement may require separate review even if its numerical weight is modest.

Worked example: Suppose throughput has weight 4 and defect prevention weight 5. Speed receives relationship scores 9 and 3, totaling 9 × 4 + 3 × 5 = 51. Temperature scores 5 and 8, totaling 5 × 4 + 8 × 5 = 60. Temperature ranks higher for investigation under these hypothetical judgments. The matrix does not reveal whether raising temperature improves or worsens either outcome, so collect directional data before setting it.

Watch the distinction: A higher weighted score is a priority for validation, not proof of causation or a recommended setting. Relationship scores are judgments; unknown direction, interactions and weak evidence can overturn the ranking.

Reference: [7]

49. Quantify measurement error

Observed variation combines process variation with measurement error. In a crossed gauge repeatability and reproducibility study, several operators measure the same representative parts repeatedly in randomized order. Repeatability is variation under the same operator and gauge; reproducibility is variation between operators under comparable conditions. Part-to-part variation should be kept separate, and operator-by-part interaction may matter if operators disagree differently on different parts. Under an additive independent-error model, variances add, so estimated process variance equals observed variance minus gauge variance. Compare gauge error with the process differences you need to detect and with the allowed tolerance, using the study's actual components. A gauge that was sufficient for a noisy process may become inadequate after improvement narrows process spread.

Worked example: Suppose observed variance is 25 square units and a gauge study estimates measurement variance at 9 square units. Estimated process variance is 25 - 9 = 16 square units, so process standard deviation is 4 units; gauge standard deviation is 3 and observed standard deviation is 5. Measurement accounts for 9/25 = 36% of observed variance under the stated model. A small apparent process difference may therefore be hard to resolve.

Watch the distinction: Subtracting standard deviations, 5 - 3 = 2, is incorrect here. Independent variance components add, so subtract squared spreads first and then take the square root of the resulting process variance.

Reference: [7]

50. Preserve time order in capability

Capability calculations describe a process distribution against specifications, but time order determines whether that distribution is predictable. Collect rational subgroups close enough in time that within-subgroup variation reflects short-term operating conditions; plot subgroup averages and ranges or standard deviations. Signals such as a point beyond a control limit or a sustained pattern suggest special causes or drift. Within-subgroup spread supports Cp and Cpk calculations, while overall spread supports Pp and Ppk. If subgroup means move over time, within spread may understate what customers experience across days or shifts. Check the measurement system, subgroup design and stability before using any index to forecast future conformance. Specification limits come from requirements; control limits are calculated from process behavior and answer a different question.

Worked example: Five readings are taken each hour. Within each hour, the readings cluster tightly, but hourly means rise through the day. Suppose the specification width is 12 units, estimated within standard deviation is 1, and overall standard deviation is 2. Then Cp = 12/(6 × 1) = 2, while Pp = 12/(6 × 2) = 1. The gap is a cue to investigate between-hour movement, not proof of future capability.

Watch the distinction: A high Cp cannot certify a drifting process. Cp uses within-subgroup spread and ignores centering; inspect the time plot and Cpk or Ppk before claiming that future units will meet specifications.

Reference: [7]

51. Plan power before testing

Power is the probability that a planned test detects a specified real effect under its assumptions. Set the smallest operationally worthwhile effect first, then estimate variation, choose the significance level and target power to size the study. Smaller effects, greater noise, stricter alpha and higher desired power generally require more observations. For two independent means with equal group size n and common standard deviation σ, the standard error of their difference is σ√(2/n). An approximate two-sided planning formula is n per group ≈ 2σ²(z for alpha/2 + z for target power)²/δ², before allowance for design complications. State assumptions about independence, variance and missing data. A nonsignificant result from a weak study does not demonstrate equivalence; an equivalence claim needs a prespecified margin and an appropriate interval or test.

Worked example: Suppose the worthwhile mean change is δ = 1 minute, σ = 6 minutes and n = 4 independent cases per group. Standard error is 6√(2/4) ≈ 4.24 minutes, making the signal-to-noise ratio only 1/4.24 ≈ 0.24. Using hypothetical two-sided alpha 0.05 and 80% power, a normal approximation gives n ≈ 2 × 36 × (1.96 + 0.84)² ≈ 565 per group. This rough result demands a feasibility review.

Watch the distinction: Failure to reject with four cases per group is not evidence that a one-minute effect is absent. The test is too noisy for that target; estimate uncertainty and plan power before interpreting a negative result.

Reference: [7]

52. Analyze paired non-normal outcomes

Before-and-after measurements on the same item or person form pairs. Calculate each after-minus-before difference and keep the pairs intact. If normality of differences is implausible, the sign test counts only positive and negative nonzero differences, testing whether positive and negative changes are equally likely under its null. It is robust to extreme magnitudes but loses their information. Wilcoxon signed-rank ranks the absolute nonzero differences and attaches their signs; its familiar location-shift interpretation relies on a reasonably symmetric difference distribution. Inspect ties, zeros and asymmetry before choosing it. Random or otherwise defensible sampling still matters, and neither test proves that the intervention caused the change when time or another condition changed too. Report the observed effect and its operational importance alongside any p-value.

Worked example: Seven paired changes contain six improvements and one decline, with no ties. Under an equal-chance sign null, the two-sided probability of at least this imbalance is 2 × [C(7,6) + C(7,7)]/2⁷ = 2 × 8/128 = 0.125. At a hypothetical 0.05 level, that is insufficient to reject the null. Wilcoxon might use magnitude information, but its assumptions and the seven actual difference sizes must be checked.

Watch the distinction: Six improvements out of seven may look decisive, but the exact two-sided sign-test p-value is 0.125. Choosing Wilcoxon solely because it may yield a smaller p-value is unjustified.

Reference: [7]

53. Interpret interacting predictors

An interaction term says the effect of one predictor changes with another predictor. In a model with coded levels A and B of -1 and +1, write the predicted response at each combination rather than reading a coefficient in isolation. For Y = β0 + βA A + βB B + βAB AB, the high-versus-low effect of B at a fixed A is 2(βB + βAB A). Its sign can reverse across A levels. Keep predictions within settings actually studied, and examine residual plots for curvature, unequal spread and unusual points. If the observations were collected rather than randomized, an interaction is an association that may still reflect confounding. When an interaction is material, describe simple effects at relevant operating settings instead of reporting only an average main effect.

Worked example: For Y = 10 + 2A + 3B - 4AB, at A = +1 the prediction is 11 for B = +1 and 13 for B = -1, so raising B lowers Y by 2. At A = -1, predictions are 15 and 1 respectively, so raising B increases Y by 14. These coded-setting contrasts show a strong interaction; check whether all four settings were actually observed and whether residuals support the model.

Watch the distinction: The +3 coefficient for B is not a universal three-unit gain. With the -4AB interaction, B's effect changes with A and even reverses direction between the coded A levels.

Reference: [7]

54. Plan an experiment that can answer its question

A useful experiment begins with the decision it must support. Define one primary response and acceptable measurement precision, then choose factor ranges that are safe and operationally meaningful. Replication repeats treatment combinations on independent experimental units and estimates noise; repeated readings of the same unit mainly assess measurement error. Randomize run order to reduce bias from time trends. Block a known nuisance source, such as material batch, by comparing treatments within each block. Decide in advance whether interactions or curvature matter, because a design that cannot estimate them cannot answer those questions. Record setup conditions, assignment and response timing so differences are attributable to treatments as far as the design allows. If the available runs are few, prioritize an interpretable comparison over a crowded design with no useful error estimate.

Worked example: Three material batches each receive one run at low temperature and one at high temperature, with the two temperatures randomized within each batch. There are six runs, and batch serves as a block. If high-minus-low responses are 2, 4 and 3 units, the average paired effect is (2 + 4 + 3)/3 = 3 units. Variation among the three differences offers limited error information; repeat runs if a precise estimate is needed.

Watch the distinction: Running all low-temperature trials first and all high-temperature trials later confounds temperature with time. Blocking by batch and randomizing within each batch protect the comparison against those nuisance effects.

Reference: [7]

55. Calculate full factorial effects

A balanced two-factor, two-level design runs all four combinations, allowing main effects and an interaction to be estimated separately. Code each factor low as -1 and high as +1. The main effect of A is the mean response at A high minus the mean at A low, averaging over B; B is calculated similarly. The AB interaction effect compares the average of same-sign corners with the average of opposite-sign corners. Equivalently, it is half the difference between A's effect at B high and at B low. With only one run per corner, the four observations fit an intercept, two main effects and interaction exactly, leaving no pure error estimate for significance testing. Replicate randomized corner runs or add suitable design points when uncertainty or curvature matters. The ICBB syllabus includes full factorial analysis.

Worked example: Responses at (A-, B-), (A+, B-), (A-, B+) and (A+, B+) are 10, 14, 12 and 20 units. A's effect is (14 + 20)/2 - (10 + 12)/2 = 6 units. B's effect is (12 + 20)/2 - (10 + 14)/2 = 4 units. The AB effect is (10 + 20)/2 - (14 + 12)/2 = 2 units. No replication means these effects lack a direct pure-error estimate.

Watch the distinction: The 20-unit response at A+ B+ does not mean setting both factors independently adds ten units. Their effects include interaction, and a single observation per corner cannot establish statistical reliability.

Reference: [7]

56. Detect curvature with center points

A two-level factorial model describes a plane across the studied factor region. At the coded center, its predicted value equals the average of balanced corner responses, because linear and interaction terms cancel. Center runs supply a direct check for departure from that prediction and, when repeated independently, an estimate of pure error at the center. A substantial center-versus-corner gap suggests curvature somewhere in the region, but does not identify which factor is curved or locate the optimum. Check residuals, time order and whether center runs were randomized among corner runs; a process drift could mimic curvature. Follow with a design that estimates quadratic terms if the decision depends on settings between corners. Keep the response objective and constraints in view, since a higher center response is favorable only when higher is desirable.

Worked example: Suppose balanced corner responses average 20 units and three independent center runs are 26, 27 and 28 units. Their mean is 27, seven units above the linear factorial prediction of 20; their sample standard deviation is 1 unit. That gap is large relative to center-run scatter and merits a curvature investigation. It does not reveal which factor causes curvature or prove that the exact center is optimal.

Watch the distinction: Adding center points does not by itself estimate every quadratic coefficient. The seven-unit discrepancy detects possible curvature; further factor-level runs are needed to map its shape and check for a process-time explanation.

Reference: [7]

57. Read fractional factorial aliases

A half-fraction of a three-factor, two-level factorial uses four rather than eight runs by choosing one factor column as a product of others. With generator C = AB, multiplying both sides by C gives the defining relation I = ABC. Multiply that relation by any effect label to find its alias: A is aliased with BC, B with AC, and C with AB. The design is resolution III, so main effects are confounded with two-factor interactions. Interpret a large apparent main effect only if the paired interaction is plausibly small or after additional data separate them. Resolution describes the shortest word in the defining relation and helps protect the effects most important to the decision. A foldover changes selected signs to break aliases, while a full factorial estimates all effects separately if resources permit.

Worked example: In a hypothetical 2^(3-1) design, set C = AB. The four runs use A and B at every low/high combination, with C fixed by their product. Since I = ABC, multiplying by A gives A = BC because A² = I. If the estimated A contrast is 8 units, it represents the combined aliased contribution of A and BC; it cannot be assigned wholly to A without an assumption or more runs.

Watch the distinction: Four runs cannot independently estimate all three main effects and all interactions. Calling an aliased A contrast a proven A effect hides the possible BC contribution created by the generator.

Reference: [7]

58. Calculate chart limits from subgroups

An X-bar chart tracks subgroup means when several continuous measurements are taken under comparable short-term conditions. Its center line is the grand mean of subgroup means. For equal subgroup size, a common range-based form places limits at grand mean ± A2 × average subgroup range, using the chart constant for that subgroup size. The companion R chart checks within-subgroup spread; interpret it before trusting the X-bar chart's variance basis. Control limits describe expected process behavior under a stable baseline, while specifications state acceptable output requirements. A point inside control limits can still represent nonconforming product, and a statistically unusual point can fall inside specifications. Preserve subgroup timing, use consistent units and investigate special causes before recalculating limits. The IASSC Black Belt guide covers the surrounding control decisions.

Worked example: Suppose equal-sized subgroups yield grand mean 50 units and average range 4 units. With a supplied A2 constant of 0.577 for that subgroup size, X-bar limits are 50 ± 0.577 × 4 = 50 ± 2.308, or 47.692 and 52.308 units. A customer specification of 48 to 52 units is narrower and serves a separate purpose. Check the R chart before interpreting the X-bar signals.

Watch the distinction: Using 48 and 52 as chart limits would confuse customer requirements with process behavior. Control limits come from subgroup data and a chart constant; specification limits come from acceptance requirements.

Reference: [7]

59. Detect small sustained shifts

A Shewhart chart responds well to a relatively large isolated departure, while CUSUM and EWMA methods accumulate information across successive observations to detect smaller sustained shifts. A CUSUM adds deviations from a target after allowing for a chosen reference value; an EWMA blends each new reading with the prior smoothed value using a chosen weight. Both need limits designed for the desired false-alarm and detection behavior. Establish a stable historical baseline and keep measurement methods consistent. Several readings on one side of the mean are a diagnostic cue, but no signal can be declared from their direction alone without the specified chart rule and parameters. When a designed rule signals, check data integrity, contain affected output if needed and seek the process change before resetting the baseline.

Worked example: Suppose six successive readings are each 0.5 unit above a stable historical mean. Their raw deviations total 6 × 0.5 = 3 units, although none breaches a hypothetical Shewhart three-sigma limit. A CUSUM would accumulate only the amount beyond its chosen reference value, so its actual signal depends on that value and its decision limit. The repeated positive direction justifies applying the prespecified small-shift rule and investigating any resulting signal.

Watch the distinction: Six readings above the mean do not automatically prove a CUSUM breach. Without its reference value, starting state and decision limit, the three-unit raw sum is insufficient to determine a signal.

Reference: [7]

60. Close the control loop

A control plan makes the improved method operational after the project team leaves. For each critical characteristic, specify the definition and measurement method, sample size and frequency, chart or other trigger, owner, immediate containment and investigation steps. Link a signal to the last known good checkpoint so potentially affected material is identified by evidence rather than guesswork. Separate process correction from product disposition: restoring a machine setting does not decide whether units made during the disturbance are acceptable. Record causes, actions and verification that performance returned to stable operation. Review leading checks and lagging defect or cost results with the process owner and finance. Include the ongoing cost of sampling, maintenance and response when judging sustained net benefit, and update the plan when the process or measurement system changes.

Worked example: A chart signals during a 100-unit lot. The owner isolates those 100 units, checks the gauge and records, traces output back to the last known good check, investigates the cause and documents each lot's disposition. Suppose verified avoided loss is $1,000 per month and monitoring costs $600 per month. Net benefit is $1,000 - $600 = $400 per month, assuming comparable volume and no omitted material costs.

Watch the distinction: A signal does not justify silently resetting chart limits or releasing the lot after a machine adjustment. Containment, measurement checks, cause verification and product disposition address different parts of the control loop.

Reference: [7]

Official references and further reading

All study guides

Choose your exam version

ICGB

ICBB

CSSGB

CSSBB

ASQ Six Sigma Green Belt (CSSGB)

1. Choose the improvement route

2. Map value and flow

3. Write a workable charter

4. Translate the customer voice

5. Set process boundaries with SIPOC

6. Plan work and decisions

7. Prioritize failure modes

8. Calculate defects and yield

9. Sample the process fairly

10. Check the measurement system

11. Separate stability from capability

12. Investigate patterns and relationships

13. Match a test to the question

14. Read experimental effects

15. Hand control to the owner

ASQ Six Sigma Black Belt (CSSBB)

16. Deploy projects against strategy

17. Value benefits over time

18. Manage adoption barriers

19. Lead a working team

20. Build a line of sight to outcomes

21. Find the true flow constraint

22. Protect data integrity

23. Handle non-normal capability

24. Choose a probability model

25. Interpret uncertainty and power

26. Compare several groups

27. Validate a regression model

28. Design a fair experiment

29. Select advanced controls

30. Design for constraints and robustness

IASSC Lean Six Sigma Green Belt (ICGB)

31. Trace the output to inputs

32. Recognize waste at the source

33. Define benefits that can be captured

34. Prioritize possible causes

35. Read distribution shape

36. Detect bias across the range

37. Study variable and attribute agreement

38. Assess attribute performance

39. Locate variation with multi-vari data

40. Use sampling distributions

41. Test normal-data questions

42. Choose a rank-based comparison

43. Test counts and proportions

44. Build and diagnose a prediction

45. Sustain a Lean improvement

IASSC Lean Six Sigma Black Belt (ICBB)

46. Verify strategic benefits

47. Expose queues in the value stream

48. Select inputs for validation

49. Quantify measurement error

50. Preserve time order in capability

51. Plan power before testing

52. Analyze paired non-normal outcomes

53. Interpret interacting predictors

54. Plan an experiment that can answer its question

55. Calculate full factorial effects

56. Detect curvature with center points

57. Read fractional factorial aliases

58. Calculate chart limits from subgroups

59. Detect small sustained shifts

60. Close the control loop

Official references and further reading

Frequently Asked Questions

Related Study Guides

Accredited Business Valuation (ABV)

Accredited Estate Planner (AEP)

Accredited Senior Appraiser (ASA - Valuation)