Measurement System Analysis & Gage R&R

Measurement System Analysis & Gage R&R

Quick Answer: Measurement System Analysis (MSA) tests whether your gage, method, and people produce data good enough to manage the process. Gage R&R decomposes measurement variation into repeatability (same person/equipment, same part) and reproducibility (different appraisers or setups). CSSGB BoK III.E (Evaluate) expects you to interpret study results—especially %GR&R, correlation of measurements, and precision—before you trust capability, control charts, or experiment outcomes.

Why MSA Belongs in Measure

Every observed data point mixes true part (process) variation with measurement-system variation:

$\sigma^2_{\text{total observed}} = \sigma^2_{\text{part}} + \sigma^2_{\text{measurement}}$

If $\sigma_{\text{measurement}}$ is large, charts will be noisy, capability indices will look worse than the process is, and DOE effects can be masked. Green Belts evaluate the measurement system early—ideally before locking baseline metrics—so improvement work targets the process, not a bad gage.

A measurement system includes the instrument, fixtures, software, procedures, environment, and appraisers. MSA is broader than “calibrate the micrometer once.”

Core MSA Concepts for Variables Data

Measurement correlation (often discussed with MSA): repeated measures of the same parts should track each other and track reference values. Weak correlation between Appraiser A and Appraiser B, or between the gage and a master, signals a reproducibility or accuracy problem even before you compute %GR&R.

Typical Crossed Gage R&R Design

A common crossed study for CSSGB-level interpretation:

• 10 parts spanning the expected process range (not only “good” parts)
• 2 or 3 appraisers who normally use the gage
• 2 or 3 trials each, parts presented in random order, appraisers blinded to prior results when possible

Software (or ANOVA / average-and-range methods) estimates variance components for parts, appraisers, equipment, and interactions. You do not need to hand-derive ANOVA tables on every exam item, but you must know what each component means and how % metrics are formed.

Calculating and Interpreting %GR&R

Define standard deviations from the variance components:

$\sigma_{\text{GRR}} = \sqrt{\sigma^2_{\text{repeatability}} + \sigma^2_{\text{reproducibility}}}$

$\sigma_{\text{TV}} = \sqrt{\sigma^2_{\text{part}} + \sigma^2_{\text{GRR}}}$

Percent GR&R relative to study (total) variation:

$\%\mathrm{GR\&R} = \frac{\sigma_{\text{GRR}}}{\sigma_{\text{TV}}} \times 100\%$

Equivalently, using variances:

$\%\mathrm{GR\&R} = \sqrt{\frac{\sigma^2_{\text{GRR}}}{\sigma^2_{\text{TV}}}} \times 100\%$

(You will also see contribution percentages that report variance shares rather than standard-deviation shares—read the label carefully.)

AIAG-style decision guidelines (widely taught for CSSGB):

Also report %Repeatability and %Reproducibility of TV to see where to improve: fixture/method noise (repeatability) vs. training/standardization (reproducibility).

Number of Distinct Categories (ndc)

ndc estimates how many groups the measurement system can distinguish across the process spread:

$\mathrm{ndc} = 1.41 \times \frac{\sigma_{\text{part}}}{\sigma_{\text{GRR}}}$

(rounded down to an integer in practice). ndc ≥ 5 is the usual minimum for using the gage to analyze or control process variation. Low ndc means the gage mostly sees its own noise, not part differences.

Precision-to-Tolerance (%P/T) Preview

When tolerance is known, %P/T compares measurement precision to the specification width (detail and criteria in the next section):

$\%\mathrm{P/T} = \frac{k \, \sigma_{\text{GRR}}}{\mathrm{USL} - \mathrm{LSL}} \times 100\%$

with $k$ commonly 6 (covering most of the measurement distribution). A system can look good on %GR&R (parts vary a lot in the study) yet poor on %P/T (tolerance is tight)—or the reverse. Evaluate both when specs exist.

Worked Example with Numbers

A digital thickness gage is studied with 10 parts, 3 appraisers, 3 trials. ANOVA software returns these standard deviations (same units as the CTQ, mm):

• $\sigma_{\text{repeatability}} = 0.012$
• $\sigma_{\text{reproducibility}} = 0.009$
• $\sigma_{\text{part}} = 0.080$

Step 1 — GRR standard deviation

$\sigma_{\text{GRR}} = \sqrt{0.012^2 + 0.009^2} = \sqrt{0.000144 + 0.000081} = \sqrt{0.000225} = 0.015$

Step 2 — Total study variation

$\sigma_{\text{TV}} = \sqrt{0.080^2 + 0.015^2} = \sqrt{0.0064 + 0.000225} = \sqrt{0.006625} \approx 0.0814$

Step 3 — %GR&R

$\%\mathrm{GR\&R} = \frac{0.015}{0.0814} \times 100\% \approx \mathbf{18.4\%}$

Interpretation: 18.4% falls in the marginal band. For a critical safety CTQ you would improve the system; for a noncritical internal metric you might accept it temporarily with documented risk while you improve fixtures or training. Repeatability (0.012) is larger than reproducibility (0.009), so equipment/method (clamping, probe placement, warm-up) is the first improvement target—not only “train the people.”

Step 4 — ndc

$\mathrm{ndc} = 1.41 \times \frac{0.080}{0.015} = 1.41 \times 5.333 \approx 7.5 \rightarrow \mathbf{7}$

ndc ≥ 5 supports using the gage for process discrimination even though %GR&R is marginal.

Step 5 — %P/T (if tolerance = 0.50 mm)

$\%\mathrm{P/T} = \frac{6 \times 0.015}{0.50} \times 100\% = \frac{0.090}{0.50} \times 100\% = \mathbf{18\%}$

Again marginal by common 10% / 30% bands—consistent with the GR&R story for this gage.

Precision vs. “Good Enough” Data

Precision is about spread of repeated measurements, not about hitting the true value on average (that is bias/accuracy). A precise but biased gage can produce tight, wrong data. MSA therefore pairs R&R with bias, linearity, and stability studies.

Correlation checks in MSA practice:

• High correlation between repeated trials → supports repeatability
• High correlation between appraisers → supports reproducibility
• High correlation with reference masters → supports accuracy (with bias near zero)

Low correlation is a red flag even if someone quotes a “passing” summary percentage from a poorly designed study.

Attribute vs. Variables MSA (Exam Awareness)

Gage R&R above is for variables (continuous) measurements. Attribute systems (pass/fail, go/no-go) use percent agreement, kappa-type association, and effectiveness metrics (true accept/reject rates)—covered with decision criteria alongside bias and P/T. Do not apply %GR&R formulas to pure go/no-go data without a variables measurement underneath.

Common Pitfalls

• Running R&R only on “good” parts → underestimates part variation → inflates %GR&R artificially.
• Appraisers know the “right” answer or measure in fixed order → underestimates reproducibility.
• Confusing % contribution based on variance with %GR&R based on standard deviation.
• Accepting a system because calibration stickers are current—calibration addresses bias/traceability, not R&R precision in use.
• Skipping MSA, then “improving” a process that was never measured well.

Bottom line for CSSGB: Evaluate measurement correlation, precision, and %GR&R with a proper parts×appraisers×trials study; use AIAG-style bands and ndc to decide whether to measure, improve the gage, or redesign the method—before you bet the project on the numbers.