Measurement systems analysis

The formula

The model behind a crossed gage R&R study: every reading is a part effect, an operator effect, their interaction, and noise:

x_ijk = μ + part_i + operator_j + (part·operator)_ij + e_ijk

Reading k of part i by operator j decomposes into who measured what, and the leftover jitter e is repeatability.

The variance ledger that comes out of the study:

σ²_total = σ²_part + σ²_GRR
σ²_GRR   = σ²_repeatability + σ²_reproducibility
σ²_reproducibility = σ²_operator + σ²_operator·part

Total observed variation splits cleanly into what the parts actually differ by and what the measurement system adds; the gage's share splits into same-operator scatter (repeatability, the equipment) and between-operator shift (reproducibility, the people and their technique) ^[2]^[3].

The headline metrics:

%GRR = 100 · σ_GRR / σ_total
%Tolerance (P/T) = 100 · 6·σ_GRR / (USL − LSL)
ndc = 1.41 · σ_part / σ_GRR        (1.41 ≈ √2)

%GRR compares the gage to the variation it's watching; P/T compares it to the spec window it's guarding; ndc counts how many distinct groups of parts the system can actually tell apart ^[1]^[3]^[4]. The quick-and-dirty range method estimates repeatability without ANOVA:

σ̂_repeatability = R̄ / d₂

Average the range across repeat readings of the same part and divide by the control-chart constant d₂ (1.128 for 2 trials, 1.693 for 3) ^[6].

Where you meet it

First article on a new gage. A supplier submitting a PPAP package to an automotive or defense prime attaches an MSA study for every gage on the control plan — it's a contract deliverable, not a nicety, and has been since QS-9000 in 1994 ^[7]^[8].
At the test stand, arguing about a noisy channel. Thrust or pressure data shows run-to-run spread and the program wants to know if the article changed or the DAQ chain did. A repeatability study on a reference load — same stimulus, repeated, across shifts — is MSA whether anyone calls it that or not ^[11].
In a capability review going sideways. Cpk came back at 1.1 and the board wants a process fix. The observed sigma includes the gage: σ_observed = √(σ²_process + σ²_gage). Until someone shows the %GRR, nobody knows how much of that 1.1 is real.
During an IATF 16949 audit. The auditor pulls a gage from the control plan and asks for the study, the acceptance criterion, and the disposition. Missing MSA records are one of the routine findings.

How it works

The standard crossed study: 10 parts spanning the process range, 3 operators, 2 or 3 trials each, parts presented blind and in random order ^[1]^[2]. Every operator measures every part every trial — that crossing is what lets ANOVA split operator effects from part effects and pull out the operator-by-part interaction, which the older average-and-range method can't see ^[2]^[3]. The interaction matters: it's the signature of an operator who reads one geometry of part differently than another — a fixturing or technique problem no amount of equipment money fixes.

All of that assumes the measurement hands the part back. A crossed study only works if every operator can remeasure every article; tensile pulls, weld break tests, single-use pyros, and hot-fire runs consume the thing being measured, and they call for a nested (destructive) design instead — each operator measures their own batch of parts, and the analysis leans on the batches being alike (Minitab: Gage R&R Study – Nested; NIST's gauge studies are built on nested designs) ^[11]^[12]. That's why the test-stand vignette above runs on a reference load: it's the one article everyone can measure again. And a go/no-go gage produces attribute data, not variables — it gets an attribute agreement analysis, not a variable gage R&R ^[13].

Acceptance runs on two gates. AIAG's %GRR bands: under 10% acceptable, 10–30% conditionally acceptable depending on the application and cost, over 30% unacceptable ^[1]. And ndc of 5 or more ^[1]^[4]. The two are the same physics viewed twice — at exactly 30% GRR, ndc works out to 1.41·√(1−0.3²)/0.3 = 4.5, right at the failing edge of both gates.

The gotchas, in the order they bite:

The percentages don't add. %GRR and %PV are ratios of standard deviations, but only variances add. A system at 30% GRR has 95.4% part variation — 125.4% total. People see the columns, assume they're shares of a pie, and misread the report. Donald Wheeler built a career-long critique on this and on the biased constants in the range method ^[5].

%GRR depends on the parts you picked. It's a ratio to total study variation. Pull ten parts off one pallet made in one hour and σ_part collapses, %GRR balloons, and a fine gage fails the study. Pick parts spanning the real process spread, or judge against the tolerance (%P/T) or a historical process sigma instead ^[1]^[4].

Ten parts is a small sample for a variance. Confidence intervals on the variance components from a 10×3×2 study are wide enough that a 25% result and a 35% result may be the same gage on different days ^[5]. Treat the bands as guidance, not physics.

A big %GRR doesn't inflate your data much — it misclassifies your parts. At 30% GRR the observed sigma only overstates the true process sigma by about 5% (1/0.954). The real damage is at the spec limits: parts near the line get passed and failed by coin flip. That's why P/T against the tolerance is the number the customer cares about.

The multiplier moved. Older studies spread the gage error over 5.15·σ (99% of a normal); the current convention and the software defaults use 6·σ (99.73%) ^[1]^[3]. Comparing a 1990s study to a new one without checking the multiplier is a 16% apples-to-oranges error.

And the classic mistake: reading a bad reproducibility number as "operator error" and starting a training program. Reproducibility absorbs fixturing, part loading, ambient shifts between shifts, and instructions vague enough that three careful people do three different things. The study points at the measurement system — the AIAG manual's title says systems, not gages, on purpose ^[1].

History

The method is a product of the automotive industry, not the statistics department. The average-and-range gage study grew up inside General Motors in the 1960s as a shop-floor technique — control-chart constants, no ANOVA, computable by hand at a bench ^[5]. It circulated on GM forms for two decades before GM handed the material to the Automotive Industry Action Group in 1989 ^[5], and AIAG — the joint quality body of GM, Ford, and Chrysler — published the first Measurement Systems Analysis reference manual in October 1990, with editions following in 1995, 2002, and 2010 ^[1].

What turned a Detroit worksheet into a global requirement was QS-9000, the Big Three's common supplier quality standard released in 1994: any shop that wanted automotive work now had to run MSA per the AIAG manual on every gage in the control plan ^[7]^[8]. QS-9000 gave way to ISO/TS 16949 and then IATF 16949, which carried the requirement forward ^[8].

The loyal opposition arrived early. Donald Wheeler and Richard Lyday's Evaluating the Measurement Process (second edition 1989, SPC Press) laid out a control-chart-based alternative — the EMP method — and Wheeler spent the following decades documenting where the AIAG arithmetic misleads: non-additive percentages, biased range constants, and variance estimates too loose to support the three-band verdicts hung on them ^[5]^[9]^[10]. NIST's own gauge-study methodology in the SEMATECH handbook runs on nested designs and pooled standard deviations, closer in spirit to EMP than to the AIAG worksheet ^[11]. The argument is still live; the AIAG bands are still what the auditor asks for.

Related tools

/tools/strain-gauge-bridge — the measurement chain most test-stand R&R studies are actually qualifying
/tools/wheatstone-bridge — the cal-bench measurement underneath it
/tools/adc-resolution — quantization sets a hard floor on repeatability; a gage can't resolve below its LSB
/tools/snr-enob — digitizer noise expressed as effective bits is a repeatability statement
/tools/loop-4-20ma — transmitter accuracy specs are what the reproducibility findings get argued against
/tools/bolt-torque — click wrenches are the floor's most-studied gage R&R subject

Sources

Written by HE in our own words from the cited sources — engineering judgment included, your stamp still required. All entries →