Gage R&RGage Repeatability & Reproducibility

Gage Repeatability and Reproducibility (Gage R&R, GR&R) is the standard Measurement System Analysis experiment for quantifying the variation contributed by the measurement system itself. The classic design: take 10 parts that span the expected range of variation, have 3 operators each measure each part 3 times in randomised order, and decompose the total variation into part-to-part (the real variation), repeatability (within-operator within-gauge variation), and reproducibility (between-operator variation).

AIAG MSA codified the technique in the 1990s; ISO 22514-7 and ASTM E2782 provide alternative international references. Gage R&R is the standard MSA approach for continuous measurements (length, weight, voltage, concentration); attribute Gage R&R (Cohen's kappa, the attribute-agreement analysis) covers go/no-go inspection.

Total variation σ²_total = σ²_part + σ²_GRR, where σ²_GRR = σ²_repeatability + σ²_reproducibility. Repeatability is the within-operator variation: same person, same gauge, same part, repeated trials — measures gauge precision under best conditions. Reproducibility is the between-operator variation: how much the answer changes when a different operator picks up the same gauge to measure the same part — measures gauge bias-by-operator interaction.

The decomposition is typically done by ANOVA (the modern AIAG method) or by the older Average and Range method. ANOVA is preferred because it separates operator-by-part interaction (a fourth variance component) which the Average and Range method lumps into reproducibility. Software (Minitab, JMP, R msa package) handles the arithmetic.

Two acceptance metrics matter. %GRR = (σ_GRR / σ_total) × 100 — the fraction of observed variation contributed by the measurement system. %P/T (precision-to-tolerance) = (6 × σ_GRR / Tolerance) × 100 — how much of the spec width the measurement system consumes.

A third metric — Number of Distinct Categories (NDC) = 1.41 × (σ_part / σ_GRR) — captures the gauge's discriminative power. NDC ≥ 5 is acceptable; NDC < 2 means the gauge effectively cannot tell parts apart. AIAG MSA expects NDC reported alongside %GRR.

• Parts — typically 10, selected to span the expected production range. Not deliberately chosen extremes (artificially inflates σ_part); not all from one lot (deflates σ_part).
• Operators — typically 3, representative of the routine operator population. Not the best 3 you have; not 3 from one shift.
• Trials — typically 3 per operator per part. More trials tighten the repeatability estimate but add cost.
• Randomisation — the 3 × 10 × 3 = 90 measurements run in random order so operators cannot anticipate the answer.
• Blinding — operators do not know which part is which during the study.
• Environmental controls — temperature, vibration, lighting consistent with routine use. Studies done in a metrology lab on a vibration-isolated bench produce optimistic numbers that don't reflect shop-floor reality.

When the measurement is attribute (go/no-go, acceptable/rejected, present/absent), the continuous-data Gage R&R doesn't apply. The attribute-agreement analysis instead asks: when 3 operators each inspect 50 parts 2 times, how often do they agree with themselves (within-operator agreement), with each other (between-operator agreement) and with the known reference (overall accuracy)?

Cohen's kappa quantifies agreement corrected for chance: kappa < 0.4 is poor agreement; 0.4–0.6 moderate; 0.6–0.8 substantial; > 0.8 excellent. For regulated inspections (visual particulate inspection, defect classification, label verification) attribute Gage R&R is the controlling MSA, and inspection programmes that have never run one routinely show kappa around 0.3 once measured.

1. Gage R&R never performed on a measurement system used for release — Cpk reported on data of unknown measurement quality.
2. Study designed with deliberately extreme parts to inflate σ_part and pass the %GRR threshold.
3. Single-trial design (1 measurement per part per operator) — repeatability not separable from reproducibility.
4. Operators chosen from one shift / one experience level — reproducibility under-estimated.
5. Calibration confused with Gage R&R — calibration confirms bias to standard, R&R quantifies precision. Both required.
6. Attribute inspections (visual particulate, label check) never assessed with attribute-agreement analysis.
7. Marginal %GRR (10–30 %) accepted without documented rationale.
8. Re-Gage R&R never performed after a gauge change, operator turnover or environmental change.

• Every analytical / measurement asset carries its Gage R&R status, study date, %GRR, %P/T, NDC and re-study cadence.
• Studies are first-class objects with the 10×3×3 design template; data entry runs from the kiosk to remove transcription error.
• %GRR / %P/T / NDC compute automatically with ANOVA; acceptance is gated against pre-defined thresholds.
• Failed Gage R&R blocks the asset from being assigned to a release measurement until remediated.
• Attribute Gage R&R templates support kappa computation for visual inspections.
• Capability calculations (Cpk, Ppk) reference the gauge's MSA status — a non-capable gauge flags every Cpk computed on its data.