CSSGB common mistakes
CSSGB · Understand the traps and how to avoid them.
Converts a Z-score or sigma level to a defect rate without applying the 1.5-sigma shift, so a 'Six Sigma' process is reported as 0.002 ppm instead of 3.4 DPMO.
Textbooks print the unshifted normal-table value and the 3.4 DPMO figure side by side, so candidates memorize them as separate facts rather than realizing one is derived from the other using the shift convention.
What to do instead
Practice the conversion chain until it is automatic: Z(long-term) = Z(short-term) - 1.5, then look up the tail area. Anchor three points on the curve: 3 sigma = 66,807 DPMO (93.32% yield), 4 sigma = 6,210 DPMO, 6 sigma = 3.4 DPMO (99.99966% yield). Drill by converting both directions.
Writes the charter problem statement as 'Line 4 is slow because the capper jams frequently,' embedding an assumed cause (and sometimes a remedy) in the description of the problem.
The team already has a pet theory by the time the charter is drafted, so causal language slips into the statement of fact. It reads naturally, which is exactly why it goes unnoticed.
What to do instead
Apply two deletion tests to every charter sentence: does it name a cause, and does it name a solution? If either is present, cut it and move it to the team's hypothesis list. Leave only what, where, when, and magnitude, quantified from baseline data.
Puts the measurable target in the business case ('save $400K by cutting rework 60%') and leaves the goal statement vague ('improve first-pass yield').
Business case and goal statement both read as justification for the project, and templates often place them next to each other, so the content migrates to the wrong box.
What to do instead
Separate the two by audience. Business case answers 'why does the organization care' in money, customer, or strategic terms. Goal statement is SMART and metric-specific, formatted as metric + baseline + target + deadline, e.g. 'Reduce solder defects from 4,200 to 1,000 DPMO by 30 June.'
Swaps repeatability and reproducibility in a Gage R&R, concluding the gauge is at fault when the variation is actually between operators.
Both terms begin with 're', both are reported as percentages of total variation, and both appear in the same ANOVA table, so they are easy to flip under time pressure.
What to do instead
Tie the definitions to the physical setup: repeatability = same appraiser measuring the same part twice = equipment variation (EV). Reproducibility = different appraisers measuring the same parts = appraiser variation (AV). Mnemonic: repeat = the same person repeats. When reading output, check which component is the larger share before writing the conclusion.
Computes and reports Cp/Cpk from data taken on a process that is visibly drifting, or reports a software-generated Ppk as the process capability.
Statistical software returns Cp, Cpk, Pp, and Ppk in one table and they differ only in which sigma is used (within-subgroup vs overall), so the distinction gets ignored and whatever number appears first is quoted.
What to do instead
Sequence the analysis correctly: chart first, capability second. Cp/Cpk require a stable process and use within-subgroup sigma; Pp/Ppk use overall sigma and describe actual observed performance. If the control chart shows instability, say the process is not capable of being assessed and go find the special cause before computing anything.
Defines alpha as the risk of missing a real defect and beta as the risk of a false alarm, then answers power questions consistently backwards.
Both are small error probabilities attached to a hypothesis test, and the producer/consumer risk framing uses language (protecting the customer) that sounds like it belongs to the other error.
What to do instead
Lock in one sentence and reuse it: alpha (Type I) = rejecting a true null = false alarm = producer's risk; beta (Type II) = failing to reject a false null = missed effect = consumer's risk; power = 1 - beta. For any question, first write which decision is being made incorrectly, then map that to the error.
Concludes that two suppliers are 'the same' or that an improvement 'made no difference' because the test returned p = 0.22.
Non-significant is intuitively read as proof of no effect, and the exam's phrasing often rewards the practical-sounding conclusion rather than the statistically correct one.
What to do instead
Rephrase every non-significant result as 'failed to detect a difference.' Then state the caveats: the test may have been underpowered, sample size may have been too small, or variation may have masked a real shift. Also separate statistical significance from practical significance before recommending an action.
Declares an improvement successful by comparing one week of post-change data with one week of pre-change data and observing that the new average is lower.
Any two short windows of a stable process differ, and the before/after average is the most obvious comparison to make, so the check feels complete without any test of whether the shift exceeds common-cause noise.
What to do instead
Before rollout, plot the pilot data and the baseline on one control chart and require the shift to show up as a signal (points beyond the baseline limits or a run-rule pattern). Where a formal test is used, run it on adequate sample sizes and report the effect size, not just the two averages.
Draws control limits at the customer's specification limits, or asks what the UCL should be set to before any data have been collected.
Spec limits are the only limits visible on the drawing or the work instruction, and it is not intuitive that a chart's limits are computed from the process's own variation rather than chosen.
What to do instead
State it as a rule: control limits = process mean +/- 3 sigma estimated from the data, and they are not specification limits. Voice of the process vs voice of the customer. Reinforce by working an example where the process improves and the control limits tighten while the spec limits stay fixed.
Marks a control chart as 'in control' whenever no point falls outside the limits, even when eight consecutive points sit on one side of the centerline.
Most chart training emphasizes the single point beyond 3 sigma, while the sensitizing rules for runs and trends are taught separately and are far less visually obvious.
What to do instead
Memorize the pattern signals that indicate a shift or trend without any point crossing a limit: 8 consecutive points on one side of the centerline, 6 or more steadily increasing or decreasing points, 2 of 3 consecutive points beyond 2 sigma, and 4 of 5 beyond 1 sigma. Scan every chart for patterns before declaring stability.