CSSC Green Belt common mistakes
CSSC Green Belt · Understand the traps and how to avoid them.
Writing the problem statement with the cause or the fix already baked in - e.g. 'Defects occur because operators were not trained on the new fixture, so we will add a training module.'
Candidates copy the sponsor's charter language or arrive with a preferred solution already chosen, so the Define phase becomes a justification exercise rather than a scoping exercise.
What to do instead
Run every problem statement through a five-part filter before leaving Define: it names a measurable gap (baseline vs target), a metric, a location/process step, a time frame, and it contains no cause and no solution. If a suspected cause appears, move it to a separate parking lot for the Analyze phase.
Stopping at a verbatim customer complaint ('customers say it's slow') and never converting it into a CTQ with an operational definition and specification limits.
VOC feels like a soft, qualitative step and candidates are eager to start measuring, so the translation chain from need to measurable characteristic is skipped.
What to do instead
Build the chain explicitly: VOC quote → stated need → CTQ characteristic → operational definition (how exactly it is measured, with what instrument, at what point) → spec limit and data source. Do not proceed to Measure until every CTQ has a written operational definition that a second person could apply identically.
Running a measurement system study with a single appraiser, or only checking repeatability (same operator, same part) while ignoring reproducibility and part-to-part variation.
One-operator studies are fast and usually 'pass', and candidates misread a repeatability-only result as a complete MSA.
What to do instead
Use a crossed Gage R&R with at least 10 parts spanning the process range, 3 operators, 3 trials, randomized order. Report %GRR of study variation (and %tolerance where relevant) plus number of distinct categories; require ndc ≥ 5 and investigate any interaction between operator and part before accepting the data.
Computing Cp/Cpk and quoting it as the process capability when the process is not in statistical control, the data are clearly non-normal, or Ppk and Cpk are used interchangeably.
Software outputs every capability index by default and candidates report the first number on the printout without checking the assumptions that make the index meaningful.
What to do instead
Sequence the check: control chart first (stable?), then distribution (normal probability plot or Anderson-Darling), then capability. Use within-subgroup sigma for Cp/Cpk and overall sigma for Pp/Ppk and state which you used; for non-normal data use a transformation or a non-normal capability method rather than forcing a normal fit.
Selecting a hypothesis test by habit or by what the software menu offers - a two-sample t-test on skewed data, a proportions test on continuous data, or repeated pairwise t-tests instead of ANOVA for three or more groups.
Candidates memorized a few tests rather than a decision logic, and under exam time pressure they reach for the most familiar tool.
What to do instead
Work a fixed decision path every time: data type (continuous vs attribute) → number of groups → paired or independent → normality and sample size → equal variances. Run pairwise tests only after a significant ANOVA/chi-square with an appropriate multiplicity adjustment.
Reading p > 0.05 as proof that there is no difference, or reading p < 0.05 as proof of a large, practically important effect - and reporting the p-value with no confidence interval or effect size.
The p-value is the most-rehearsed number in the course, so it gets treated as the answer instead of as one input, and power/sample size is rarely considered.
What to do instead
Always report the confidence interval and the effect size alongside p, state the practical significance threshold set in Define, and phrase conclusions as 'insufficient evidence of a difference given this sample size' rather than 'no difference'. Mention power or sample size when a test is non-significant.
Changing several process inputs at once on the line and attributing the resulting shift to whichever factor the team believed in, with no design structure, randomization, or blocking.
It feels faster than designing an experiment, and the team wants visible action; without randomization the effects are confounded with time, shift, and operator.
What to do instead
Screen with a designed experiment: fractional factorial for many factors, full factorial for a few, always with randomization, blocking on known nuisance variables (shift, machine, lot), and center points to test for curvature. Confirm the identified factors in a verification run before institutionalizing the setting.
Declaring the improvement successful from a before/after bar chart of two averages, without a pilot, without a statistical comparison, and without checking that the gain persists beyond the trial period.
The team has invested effort and wants closure; the Improve tollgate gets treated as a celebration rather than a verification step.
What to do instead
Pilot on a defined scope, compare pilot vs baseline with an appropriate test or a control chart rule (a sustained shift, not a single point), confirm the CTQ metric moved by at least the target amount, and re-check after a defined sustain period before closing Improve.
Building an SPC chart with limits calculated from a handful of points, from subgroups that mix different streams, or recalculating limits every time a point falls outside them.
Candidates want the chart to look clean, and they confuse control limits with specification limits or with the latest data's average.
What to do instead
Establish limits from 20–25 rational subgroups of homogeneous, in-control data, freeze them, and only revise after a documented, assignable cause plus a process change. State explicitly that control limits come from the process, not the customer, and that subgrouping must reflect the rational subgroup (within-batch, not across streams).
Writing a control plan whose only control is 'retrain operators' or 'monitor the metric', with no named owner, no reaction plan for an out-of-control signal, and no audit frequency.
Training and monitoring are the easiest controls to write down and require no engineering, so they are used as placeholders that satisfy the form but not the process.
What to do instead
For each X and each Y, specify: the control method (error-proofing preferred, then automation, then SPC, then procedure), the sample size and frequency, the responsible role by name, the specific reaction plan when the signal fires, and the audit interval. Anything written as 'training' must be paired with a verifiable check that the behaviour persists.