Lean Six Sigma Green Belt — Complete Study Guide (2026)

Turning intentions into trustworthy numbers

Measure converts Define's intentions into numbers you can defend. It establishes how the process performs today, using data whose integrity you have verified. The Green Belt's work here is fourfold: classify the data being collected, map the process as it truly runs, prove the measurement system itself is reliable, and compute the defect and capability metrics that fix the baseline. A wrong baseline sinks the entire project — so this phase is as much about the integrity of the measurement as about the measurement itself.

Data types: continuous vs. attribute, and the four scales

The single most tested distinction in Measure is continuous (variable) versus attribute (discrete) data.

• Continuous data is measured on a scale that can, in principle, take any value between two points — length in centimeters, time in seconds, temperature, weight.
• Attribute data is counted or categorized — pass/fail, number of scratches, defect type.

Continuous data is preferred for analysis because it carries far more information per observation: it reveals process behavior and detects change with much smaller sample sizes than attribute data. If a question offers you the choice, continuous wins on information density.

Beneath this sits the nominal → ordinal → interval → ratio hierarchy:

• Nominal — labels with no order (defect type, machine A/B/C).
• Ordinal — ranked, but with unequal or undefined gaps (poor / fair / good / excellent).
• Interval — equal spacing but no true zero, so ratios are meaningless (°C: 20° is not "twice as hot" as 10°).
• Ratio — equal spacing and a true zero, so ratios hold (length in cm: 20 cm genuinely is twice 10 cm).

Knowing the scale dictates which statistics and charts are legitimate.

Process mapping: the right map for the question

Before measuring, document how the process actually runs. Each map answers a different question:

• A detailed flowchart shows the real sequence of steps and decisions, exposing rework loops, redundant checks, and hidden hand-offs. In standard notation a diamond is a decision and a rectangle is an activity; ovals are start/end.
• A swim-lane (deployment) flowchart adds a column per person/department, making hand-offs and ownership gaps visible — useful when the problem lives in the transitions between groups.
• A value stream map (VSM) goes beyond a flowchart by overlaying data — cycle times, wait times, inventory, and the information flow that triggers each step — so the team can separate value-added from non-value-added time and compute lead time.
• A spaghetti diagram traces the physical travel path of a person, part, or document, revealing wasted motion and transport.

Sampling without measuring everything

Measuring an entire population is usually impractical, so teams sample.

• Simple random sampling gives every item an equal and independent chance of selection, minimizing selection bias.
• Stratified sampling divides the population into meaningful groups (shifts, machines, regions) and samples within each, guaranteeing representation of every group.
• Systematic sampling takes every k-th item — efficient, but dangerous if the process has a cycle that lines up with the interval.

A related idea is the rational subgroup: samples collected close together in time/condition so that within-subgroup variation reflects only common cause, and special-cause shifts appear between subgroups. This principle underlies all control-chart subgrouping. Remember the vocabulary contrast: a statistic (like the sample mean x̄) summarizes a sample; a parameter (like μ or σ) describes the whole population. And larger samples shrink the standard error (which scales with 1 ÷ √n), raising precision and statistical power — a bigger n does not create normality or fix bias.

Measurement System Analysis and Gage R&R

Every observed value is the true value plus measurement error, so before trusting any baseline you must prove the measurement system is adequate. That is Measurement System Analysis (MSA).

For continuous data the workhorse is Gage Repeatability and Reproducibility (Gage R&R), which partitions measurement variation into two parts:

• Repeatability — variation when one appraiser measures the same item multiple times with the same gage (the equipment's own inconsistency).
• Reproducibility — variation between different appraisers measuring the same items (the human/method inconsistency).

A common rule of thumb reads the percentage of total variation consumed by measurement error (%GRR):

• under 10% — acceptable,
• 10–30% — marginal (accept depending on cost, criticality, and application),
• over 30% — unacceptable; the measurement system must be fixed before the data can be used.

A related output is the number of distinct categories (ndc), which should be ≥ 5 for the system to resolve enough different part levels to be useful.

Beyond R&R, a good system must also be accurate (low bias — the gap between the measured average and a reference value), linear (bias stays constant across the range), and stable (unchanging over time). Distinguish accuracy (closeness to the true value) from precision (tightness of repeated readings) — a gage can be precise yet inaccurate. For attribute (pass/fail) inspection, the right MSA is an attribute agreement analysis, which checks whether appraisers agree with each other and with a known standard. Underlying all of this is the operational definition — a precise, agreed way to measure a characteristic so two people record the same thing.

Defect metrics: DPU, DPO, and DPMO — worked

These metrics let processes of different complexity be compared on one scale. First, a vocabulary point the exam tests: a defect is any single instance of nonconformance, while a defective is a unit with one or more defects. One defective unit can carry several defects.

The three formulas:

• DPU (Defects Per Unit) = total defects ÷ total units
• DPO (Defects Per Opportunity) = defects ÷ (units × opportunities per unit)
• DPMO (Defects Per Million Opportunities) = DPO × 1,000,000

Worked example. Inspect 500 units, each with 8 opportunities for a defect, and find 20 defects.

1. Total opportunities = 500 × 8 = 4,000
2. DPO = 20 ÷ 4,000 = 0.005
3. DPMO = 0.005 × 1,000,000 = 5,000

An "opportunity" is any characteristic or chance on a unit where a defect could occur, and the opportunity count is a modeling choice the team must define consistently — inflating opportunities artificially lowers DPMO, which is how teams flatter their sigma level. Keep the definition disciplined.

Yield, sigma level, and the 1.5σ shift

Yield is the fraction of output that is good.

• First-pass (first-time) yield (FPY) counts only units that pass a step without any rework — this exposes the "hidden factory" of rework that final-inspection yield conceals.
• Rolled Throughput Yield (RTY) multiplies the first-pass yields of every step: RTY = Y₁ × Y₂ × … × Yₙ.

Worked example. A four-step process, each step at 0.95 first-pass yield: RTY = 0.95 × 0.95 × 0.95 × 0.95 = 0.95⁴ ≈ 0.815 (81.5%) — far below the 0.95 any single step reports, because losses compound. A useful companion: for a Poisson defect process, a step's throughput yield ≈ e^(−DPU) (e.g., DPU = 0.05 → yield = e^(−0.05) ≈ 0.951), and normalized yield = RTY^(1/k) gives the geometric-average per-step yield across k steps.

Process sigma expresses how many standard deviations fit between the process mean and the nearest specification limit — more sigmas, fewer defects. By convention, Six Sigma quotes long-term sigma with a 1.5σ shift built in to account for drift over time, which is why the landmark conversions are:

Memorize the two anchors the exam leans on: 6σ = 3.4 DPMO and 3σ ≈ 66,807 DPMO (~93.3% yield). Questions convert both directions using this table.

A first look at capability: Cp and Cpk

Capability is developed fully in the Control chapter, but the baseline is set here, so know the two indices:

• Cp = (USL − LSL) ÷ (6σ) — potential capability; whether the spread could fit the tolerance, ignoring centering.
• Cpk = min[(USL − mean) ÷ 3σ, (mean − LSL) ÷ 3σ] — accounts for centering by taking the worse of the two one-sided ratios.

Worked example. USL = 60, LSL = 40, σ = 2: Cp = (60 − 40) ÷ (6 × 2) = 20 ÷ 12 = 1.67.

Establishing baseline capability (baseline Cpk or sigma) at the end of Measure gives the "before" benchmark against which Improve-phase gains are measured.

Common traps

• Choosing attribute data when continuous is available. Continuous carries more information and needs smaller samples.
• Treating an interval scale (°C) as if ratios were meaningful, or forgetting that a ratio scale needs a true zero.
• Swapping repeatability (same appraiser, same part) and reproducibility (different appraisers). "Reproduce" = across people.
• Reading a %GRR of 22% as unacceptable — it's marginal (10–30%). Only > 30% is unacceptable.
• Forgetting to multiply by the opportunity count or by 1,000,000 in a DPMO problem.
• Averaging step yields instead of multiplying them for RTY. RTY is always lower than any single step.
• Confusing a defect (one nonconformance) with a defective (a unit with ≥ 1 defect).

What the exam tests here

Expect: continuous vs. attribute data and the four measurement scales; matching a process map (flowchart / swim-lane / VSM / spaghetti) to a need; sampling methods and rational subgrouping; repeatability vs. reproducibility and the %GRR acceptance bands (< 10 / 10–30 / > 30), plus ndc ≥ 5 and attribute agreement analysis; DPU/DPO/DPMO calculations (very common); FPY vs. RTY and the RTY multiplication; and DPMO-to-sigma conversion using the 1.5σ-shift landmarks (especially 3.4 = 6σ and ~66,807 = 3σ).