Six Sigma Green Belt / Black Belt — ASQ

Six Sigma Green Belt / Black Belt — ASQ

Credential header

• Credential: Certified Six Sigma Green Belt (CSSGB) / Certified Six Sigma Black Belt (CSSBB)
• Issuing body: American Society for Quality (ASQ); ANAB-accredited under ISO/IEC 17024
• Jurisdiction: Global (US-strong); ASQ's is the most recognized Six Sigma credential in US job postings
• Level: Green Belt — practitioner (works on improvement teams, typically part-time on Six Sigma); Black Belt — advanced practitioner (leads complex projects, mentors Green Belts)
• Prerequisites: CSSGB: 3 years of full-time paid work in the Green Belt Body of Knowledge. CSSBB: 3 years in the Black Belt Body of Knowledge plus at least one completed Six Sigma project with a signed sponsor affidavit (or two completed projects in lieu of the experience) — without the affidavit you cannot sit the exam. Also offered: Master Black Belt (CMBB).
• Renewal: 18 recertification units (RUs) every 3 years, or retake the exam (per ASQ's official recertification policy). ⚠ One 2026 industry analysis claims ASQ's current fact sheet lists CSSGB as a lifetime certification with no recertification — this conflicts with standing ASQ policy and could not be confirmed on asq.org; verify renewal status on asq.org before relying on it.

Exam format

Both exams are open-book (one bound reference plus calculator — a meaningfully different experience from closed-book exams; tab your reference ruthlessly). Computer-delivered CSSBB: 165 questions (150 scored, 15 unscored), 4.5-hour appointment with 4 hours 18 minutes of exam time, English only. CSSGB computer-delivered: 110 questions (100 scored). ASQ does not publish a fixed passing score; scoring is equated (reported ~550/750 scale on older paper forms). The Black Belt Body of Knowledge adds enterprise-wide deployment, advanced DOE, DFSS, and deeper statistics to the Green Belt base.

DMAIC and DMADV

DMAIC (improve an existing process): Define → Measure → Analyze → Improve → Control.

• Define: project charter, VOC (voice of the customer), CTQs (critical-to-quality characteristics), SIPOC (Suppliers–Inputs–Process–Outputs–Customers), business case. Tollgate question: are we solving the right problem for the right customer?
• Measure: operational definitions, data collection plan, measurement system analysis (GR&R), baseline capability (Cp/Cpk, DPMO, sigma level). Tollgate: can we trust our numbers?
• Analyze: exploratory data analysis, hypothesis testing, regression, ANOVA, root-cause analysis (5 Whys, fishbone). Tollgate: what actually drives the defect?
• Improve: DOE, pilot solutions, cost-benefit, FMEA on the new process. Tollgate: does the fix work and is it worth it?
• Control: control plan, SPC, standardized work, documented handover to the process owner. Tollgate: will the gain hold?

DMADV (Design for Six Sigma — create a new process/product when no process exists or DMAIC can't reach the target): Define → Measure → Analyze → Design → Verify. Use DMADV when the process doesn't exist or incremental improvement is mathematically insufficient; use DMAIC when a process exists and can be fixed. Exam trap: choosing DMAIC for a from-scratch design problem.

Roles: Champion (executive sponsor, owns the business case), Master Black Belt (coaches Black Belts, owns the program), Black Belt (full-time project leader), Green Belt (part-time practitioner), Yellow Belt (team member, basic tools).

Capability indices: Cp, Cpk, Pp, Ppk — derived, not memorized

Setup: a characteristic has specification limits LSL, USL (from the customer), and the process produces output with mean μ and standard deviation σ.

Potential capability (Cp) asks: if the process were perfectly centered, how many times does the process spread fit inside the spec width? The "process spread" is ±3σ (6σ total — the range containing 99.73% of a normal distribution).

Cp = (USL − LSL) / (6σ), dimensionless.

Actual capability (Cpk) asks: given where the process actually sits, what is the worst-side margin? Compute capability for each side and take the minimum:

Cpk = min[(USL − μ)/(3σ), (μ − LSL)/(3σ)], dimensionless.

Why the minimum: the process fails on whichever side is closer to the mean. Cpk ≤ Cp always, with equality only when perfectly centered (μ at the midpoint). A Cpk of 1.33 means the nearer spec limit is 4σ from the mean; 2.0 means 6σ — the origin of "Six Sigma" (3.4 DPMO at 6σ assumes the conventional 1.5σ long-term shift).

Cp vs Pp (and Cpk vs Ppk): Cp/Cpk use σ estimated within subgroups (short-term, from control-chart data, e.g., R-bar/d2) — "what the process could do." Pp/Ppk use the overall sample standard deviation s (long-term, all data) — "what the process actually did." Pp ≤ Cp whenever the process drifts between subgroups. Exam trap: a stable process has Cp ≈ Pp; a big gap signals instability or drift, and you must stabilize (control charts) before quoting capability.

Assumptions: normality (check it — transform or use non-normal capability methods if violated), statistical control (never quote capability for an out-of-control process), and measurement error small relative to process variation (verified by MSA).

Hypothesis testing: the decision framework

Every test follows the same logic: H0 (null: no effect/no difference) vs H1 (alternative); choose α (usually 0.05, the Type I error rate — rejecting a true H0); compute the test statistic and p-value; reject H0 if p < α. Type II error (β) is failing to reject a false H0; power = 1 − β.

• 1-sample t-test: is the mean different from a target? t = (x̄ − μ0) / (s/√n), df = n−1.
• 2-sample t-test: do two process means differ? (Check equal-variance assumption; use Welch's variant otherwise.)
• Paired t-test: before/after on the same units — more powerful because pairing removes unit-to-unit variation. Use it whenever the data are naturally paired.
• ANOVA: do 3+ group means differ? F = between-group variance / within-group variance. A significant F only says some difference exists — follow with Tukey pairwise comparisons.
• Chi-square test: association between categorical variables, χ² = Σ (O − E)²/E.
• Test of proportions: defect-rate comparisons (z-test for two proportions).
• Practical vs statistical significance: with n = 10,000, a 0.1% shift is "significant" but meaningless. Always pair the p-value with an effect size and a confidence interval.

Control charts: which chart, and the limits

Control charts separate common-cause variation (inherent, stable) from special-cause variation (assignable, investigate). A point outside the control limits, or non-random patterns (runs, trends — Western Electric rules), signals special cause.

Choosing the chart:

• Continuous data, subgroups of 2–10: X-bar and R chart (X-bar tracks the mean, R tracks within-subgroup spread).
• Continuous data, individual readings (n=1): I-MR chart (individuals and moving range).
• Attribute data, constant sample size, fraction defective: p chart; count defective: np chart.
• Attribute data, defects per unit with constant area: c chart; varying area: u chart.

Control limit formulas (3-sigma limits):

• X-bar chart: UCL/LCL = x̄̄ ± A2·R̄ (A2 from standard tables by subgroup size; for n=5, A2 = 0.577).
• R chart: UCL = D4·R̄, LCL = D3·R̄ (n=5: D4 = 2.114, D3 = 0).
• p chart: UCL/LCL = p̄ ± 3·√[p̄(1−p̄)/n].
• c chart: UCL/LCL = c̄ ± 3·√c̄.

Why 3 sigma: ±3σ limits give a false-alarm rate of ~0.27% per point for normal data — rare enough that an out-of-limit point is worth investigating, frequent enough across many points that you need the run rules too. Control limits ≠ specification limits — the most tested confusion on the exam. Control limits come from the process (voice of the process); spec limits come from the customer (voice of the customer). A process can be in control and still produce 100% defects.

DOE basics and regression

Full factorial DOE: with k factors at 2 levels each, run 2^k combinations. A 2² design (factors A, B; levels −, +) gives 4 runs; the main effect of A = (average response at A+) − (average response at A−); the interaction AB = half the difference between A's effect at B+ and at B−. Randomize run order (defeats lurking time trends), replicate (estimates pure error), block (isolates known nuisance variation). Fractional factorials (2^(k−p)) trade runs for aliased interactions — resolution IV keeps main effects clear of two-factor interactions.

Regression: simple linear model y = β0 + β1·x + ε. R² = fraction of variation explained (1 − SS_res/SS_tot) — high R² with a curved residual plot means the model is wrong, not good. Check residuals for normality, constant variance, and independence before trusting any p-value. Multiple regression adds predictors; watch for multicollinearity (VIF). Correlation ≠ causation — DOE, not regression on observational data, establishes cause.

MSA (GR&R): %GR&R = 100 × (σ_measurement / σ_total). Under 10% acceptable, 10–30% conditional, over 30% unacceptable. Always do MSA in Measure phase before believing baseline data.

Worked examples

Example 1 — Capability indices. A shaft diameter has USL = 25.40 mm, LSL = 25.00 mm. Process data: μ = 25.28 mm, within-subgroup σ = 0.060 mm, overall s = 0.075 mm.

Cp = (25.40 − 25.00)/(6 × 0.060) = 0.40/0.36 = 1.11. Cpk = min[(25.40−25.28)/(3×0.060), (25.28−25.00)/(3×0.060)] = min[0.12/0.18, 0.28/0.18] = min[0.667, 1.556] = 0.67. Interpretation: potential capability is marginal (1.11) and actual is poor (0.67) — the process is off-center toward USL. Pp = 0.40/(6×0.075) = 0.89 < Cp = 1.11, indicating some between-subgroup drift. Action: center the process first (cheapest fix), then reduce variation.

Example 2 — X-bar and R limits. Twenty-five subgroups of n = 5 give x̄̄ = 50.02 mm, R̄ = 0.18 mm. For n = 5: A2 = 0.577, D4 = 2.114, D3 = 0.

X-bar limits: 50.02 ± 0.577 × 0.18 = 50.02 ± 0.104 → UCL = 50.124 mm, LCL = 49.916 mm. R chart: UCL = 2.114 × 0.18 = 0.381 mm, LCL = 0 × 0.18 = 0 mm. A subgroup mean of 50.15 mm would plot above the X-bar UCL — investigate for special cause (tool wear, material lot change) rather than adjusting the process blindly.

Example 3 — Two-sample t-test. Old process: n1 = 30, x̄1 = 12.4 min, s1 = 1.8 min. New process: n2 = 30, x̄2 = 11.1 min, s2 = 1.6 min. Is the new process faster (α = 0.05)?

Pooled s ≈ √[(1.8² + 1.6²)/2] = √(3.24+2.56)/2 = √2.9 = 1.703 min. Standard error of difference = 1.703 × √(1/30 + 1/30) = 1.703 × 0.2582 = 0.4397 min. t = (12.4 − 11.1)/0.4397 = 1.3/0.4397 = 2.956, df = 58. Critical t (one-sided, 0.05) ≈ 1.672. Since 2.956 > 1.672, reject H0 — the new process is statistically significantly faster. 95% CI for the difference: 1.3 ± 2.00×0.4397 = 1.3 ± 0.879 → (0.42, 2.18) min. Practical check: a ~1.3-minute saving per cycle — multiply by annual volume to judge whether it justifies the changeover cost.

Example 4 — 2² DOE. Factors: temperature (A: 150°C/180°C), pressure (B: 2 bar/4 bar). Yields: (150,2) = 72%, (180,2) = 78%, (150,4) = 74%, (180,4) = 88%.

Main effect A = (78+88)/2 − (72+74)/2 = 83 − 73 = 10 percentage points. Main effect B = (74+88)/2 − (72+78)/2 = 81 − 75 = 6 points. Interaction AB = [(88−74) − (78−72)]/2 = (14 − 6)/2 = 4 points — temperature's effect is stronger at high pressure. Recommended setting: A high, B high (88% observed), then confirm with a confirmation run before changing the standard.

Example 5 — DPMO to sigma level. A process produces 12,000 units with 45 defects, each unit having 3 opportunities for a defect. DPMO = 45 / (12,000 × 3) × 1,000,000 = 45/36,000 × 1,000,000 = 1,250. Yield = 1 − 1,250/1,000,000 = 0.99875. The (short-term) sigma level ≈ NORMSINV(0.99875) + 1.5 ≈ 3.0 + 1.5 = 4.5σ. (The +1.5 is the conventional long-term shift; quote it explicitly when converting.)

Standards cross-reference

• ASQ CSSGB / CSSBB Bodies of Knowledge — the exam syllabi; study the current BoK, not a generic Six Sigma book.
• ISO 13053 (quantitative methods in process improvement — Six Sigma) — the international standard behind DMAIC practice.
• ISO 18404 (competency for Six Sigma personnel) — role definitions.
• AIAG SPC / MSA manuals — the automotive-origin reference for control charts and gauge studies, widely used beyond automotive.
• ISO/IEC 17024 — the accreditation standard for ASQ's personnel certification itself.

Pitfalls and exam strategy

1. Open-book ≠ easy. You get ~1.6 minutes per question on the Black Belt exam. If you're deriving the ANOVA table from scratch during the exam, you've already lost — tab your reference by BoK area and practice retrieving formulas in under 30 seconds.
2. Capability before control is backwards. Any question asking for Cp/Cpk on an out-of-control process: the answer is "bring it into control first" — never quote capability from unstable data.
3. Control limits vs spec limits is the single most tested confusion. Repeat: control limits come from the process, spec limits from the customer.
4. Paired vs two-sample. Before/after on the same units → paired t-test. The exam loves giving paired data to candidates who reach for the two-sample test.
5. Cpk ≤ Cp always. If your arithmetic gives Cpk > Cp, you made an error — use it as a built-in check.
6. Don't confuse DMADV with DMAIC. New process or target unreachable by improvement → DMADV/DFSS.
7. The affidavit is a gate, not a formality. Start documenting your Black Belt project (charter, data, sponsor sign-off) before you start studying — candidates fail eligibility more often than the exam.

Practice problems

1. USL = 10.50, LSL = 9.50, μ = 10.10, σ_within = 0.12. Compute Cp and Cpk. Is the process capable (target Cpk ≥ 1.33)?
2. For subgroups of n = 4, x̄̄ = 20.00 g, R̄ = 0.60 g. Compute X-bar chart control limits (A2 = 0.729 for n = 4).
3. A p-chart has p̄ = 0.04 with n = 200 per subgroup. Compute the control limits. A subgroup shows p = 0.075 — in or out of control?
4. Before/after cycle times (minutes) on the same 8 orders: before (22, 25, 21, 28, 24, 26, 23, 27), after (20, 22, 20, 24, 21, 23, 21, 24). Which test applies, and is the improvement significant at α = 0.05? (Compute the paired differences and t-statistic.)
5. A 2² DOE on cure time gives: A−B− = 45, A+B− = 52, A−B+ = 48, A+B+ = 61 (tensile strength, MPa). Compute both main effects and the interaction. Which setting maximizes strength?
6. 8,000 invoices, 96 billing errors, 2 opportunities per invoice. Compute DPMO.

Answers: 1. Cp = 1.00/(6×0.12) = 1.389; Cpk = min[(10.50−10.10)/0.36, (10.10−9.50)/0.36] = min[1.111, 1.667] = 1.11 < 1.33 — not capable; center the process (mean is above midpoint 10.00). 2. 20.00 ± 0.729×0.60 = 20.00 ± 0.437 → UCL = 20.437 g, LCL = 19.563 g. 3. σ_p = √[0.04×0.96/200] = √0.000192 = 0.01386; limits = 0.04 ± 3×0.01386 = 0.04 ± 0.0416 → UCL = 0.0816, LCL = max(0, −0.0016) = 0. p = 0.075 < 0.0816 → in control (but watch the upward drift). 4. Paired t-test. Differences (before−after): 2, 3, 1, 4, 3, 3, 2, 3; mean d̄ = 2.625, s_d = 0.916; t = 2.625/(0.916/√8) = 2.625/0.324 = 8.10, df = 7, critical (one-sided 0.05) = 1.895 → highly significant improvement. 5. Effect A = (52+61)/2 − (45+48)/2 = 56.5 − 46.5 = 10 MPa; effect B = (48+61)/2 − (45+52)/2 = 54.5 − 48.5 = 6 MPa; AB = [(61−48) − (52−45)]/2 = (13−7)/2 = 3 MPa; best setting A+B+ = 61 MPa. 6. DPMO = 96/(8,000×2) × 1,000,000 = 6,000.

Original ETS reference writing, verified against public sources on 2026-09-30. Thin sourcing is flagged inline. Not legal advice — confirm current editions with the issuing body.