Measure: Trust the Data Before You Analyze It
What Measure must produce
Measure establishes the baseline and proves the measurement system can see the variation you care about. Core concepts: operational definitions so two people measure the same thing identically; data type - continuous (time, cost, length) versus discrete/attribute (pass-fail, counts); measurement system analysis (MSA) before routine data collection - Gage R&R where repeatability is equipment variation and reproducibility is appraiser variation (rough guide: %GRR under 10% acceptable, 10–30% conditional, above 30% unacceptable), plus bias, linearity, stability, and attribute agreement analysis for discrete judgements; sampling approaches (random, stratified, systematic, rational subgrouping); sample size driven by confidence, power and the size of the effect you must detect; baseline capability using Cp/Cpk (within-subgroup, short-term) and Pp/Ppk (overall, long-term), DPMO and sigma level; a detailed process map and a written data collection plan.
Traps questions exploit
- Running capability before MSA, or capability on an unstable process - indices are meaningless until the process is in control.
- Swapping Cp (spread only) with Cpk (spread plus centering); Cpk is always ≤ Cp.
- Confusing accuracy/bias (on target) with precision (tight spread).
- Treating attribute counts as if they were continuous measurements with a normal distribution.
Worked example
USL = 10.0, LSL = 0.0, mean = 6.5, within-subgroup sigma = 1.0. Cpk = min[(10 − 6.5)/3, (6.5 − 0)/3] = min(1.17, 2.17) = 1.17. Rule of thumb: Cpk ≥ 1.33 for an existing process, ≥ 1.67 for a new or safety-critical one.