SPCVisualsTwo things this page does, in order. First, how to actually run a crossed ANOVA Gage R&R in Minitab, properly, including the option that quietly changes your answer. Second, what to do when the person who needs the study does not have a seat.
If you have Minitab open, part one is the part you want.
Running a crossed Gage R&R in Minitab
Stack your data first
Minitab wants the study stacked — three columns, one row per measurement:
C1 Part C2 Operator C3 Measurement
1 A 0.612
1 A 0.610
1 B 0.615
2 A 0.598
...
Not a matrix with operators across the top. This is where most of the time goes,
because the gauge study came off a paper form or a spreadsheet laid out for
humans. Data → Stack → Blocks of Columns will do it if you have a wide layout.
A standard AIAG crossed study is 10 parts × 3 operators × 3 trials = 90 rows. Fewer parts than 10 and the ndc figure gets unreliable fast.
The menu path
Stat → Quality Tools → Gage Study → Gage R&R Study (Crossed)
Fill in Part numbers, Operators and Measurement data, then choose the
method:
- ANOVA — what you want. It separates repeatability from reproducibility and tests whether the part-by-operator interaction is real.
- Xbar and R — the older range method. Faster to do by hand, and it cannot detect an interaction at all. It exists for continuity with paper forms.
Use ANOVA unless a customer's format specifically demands the range method.
The option that changes your answer
Click Options. Two fields matter:
Process tolerance — enter the spec width (USL − LSL), or the one-sided
distance if the characteristic is one-sided. Leave it blank and Minitab cannot
compute %Tolerance, which is usually the number the customer actually asked
for. Blank tolerance is the most common reason a study comes back missing the
column somebody wanted.
Alpha to remove interaction term — default 0.05. This is the one worth
understanding. Minitab first fits the full model with the part × operator
interaction. If that term's p-value is above alpha, Minitab drops it and
refits without it, and the reduced model is what your %GRR is computed from.
That is a defensible default, and it means the model you reported may not be the model you think you ran. Read the session output: it prints both ANOVA tables when it drops the term, and says so. A real interaction — operators measuring big parts differently from small ones — is a finding about the gauge and the operators' technique, not a nuisance parameter. Do not let it disappear silently.
Study variation is the other field: 6 standard deviations by default in
current versions, 5.15 in older ones and in some customer formats. It scales
%Study Var, so a study that "fails" under one and "passes" under the other has
not changed at all. Check which one the customer's form assumes.
Reading the output
The Gage R&R table gives you two percentage columns and they answer different questions:
- %Contribution is variance-based. The columns add to 100. Use it to see where the variation lives.
- %Study Var is standard-deviation-based. The columns do not add to 100. This is the one AIAG bands are written against.
The bands:
| %GRR (of study variation or tolerance) | Verdict |
|---|---|
| Under 10% | Acceptable |
| 10–30% | Conditional — acceptable depending on application cost and risk |
| Over 30% | Not acceptable — fix the measurement system |
And ndc (number of distinct categories) should be 5 or more. Below that, the gauge cannot reliably tell your parts apart, whatever the percentages say. Minitab truncates ndc rather than rounding.
Six graphs come out alongside. The two worth looking at first are the R chart by operator — out-of-control points there mean an operator is measuring inconsistently, and the study is compromised before you read a single percentage — and measurement by part, which should show clear separation between parts. If it does not, the parts you chose do not span enough of the process range, and %GRR will look terrible for a reason that has nothing to do with the gauge.
That last point deserves emphasis: parts for a Gage R&R must be selected to represent the process spread, not pulled as ten good ones. Ten parts from the middle of the distribution make a perfectly good gauge fail.
When you do not have a licence in front of you
The maths above is not proprietary. It is a two-way ANOVA with an interaction term and a handful of variance-component ratios, specified publicly in the AIAG MSA manual.
Our Gage R&R calculator runs it in a browser. Paste the same three columns — part, operator, reading — and it returns:
- The full two-way ANOVA table, with the part × operator interaction tested rather than assumed away
- Repeatability and reproducibility separated
- %Contribution, %Study Var and %Tolerance
- ndc, against the AIAG bands above
Free, no signup, and nothing is stored — the data goes to the calculation and comes back with the answer.
The honest case for it is narrow and specific: the study is on the shop floor and the licence is on someone's desk. A supplier sends a study to check. A production engineer wants to know whether a gauge is worth arguing about before booking a quality engineer's afternoon. A customer's PPAP pack arrives and the %GRR looks wrong. Those are all real, and none of them justify a seat.
Where Minitab is straightforwardly better
Said plainly, because finding out halfway through a PPAP is worse:
- Nested Gage R&R for destructive testing, where no two operators can measure the same part. We do crossed studies only.
- Expanded Gage R&R for additional factors — multiple gauges, sites, setups.
- Attribute Agreement Analysis for go/no-go and visual inspection, with Fleiss' kappa. Different analysis entirely; we do not have it.
- Gage linearity and bias, and gage run charts.
- Everything else in the box — DOE, regression, reliability, the full hypothesis-testing suite. See our Minitab comparison for the honest version of that argument.
- A validated package in a regulated environment. If your quality system names Minitab, use Minitab. Revalidation costs more than the licence.
If your MSA work is nested, expanded, attribute, or part of a validated system, this page has told you to buy Minitab, and it means it.
Where we are the better shape
The gauge study is a one-off. What happens after it is not, and that is the part a statistics package is not built for.
A gauge that passed in March is a gauge that passed in March. The thing that tells you it has drifted since is a chart that updates itself as the line runs, on a screen someone can see, with limits frozen from a baseline and rules that trip without anyone looking. That is what SPC Visuals is: charts priced per production line, in a browser, fed by CSV or REST API or an on-prem collector.
Most plants need both. A statistics package on the quality engineer's desk for studies, live charts on the floor for control. They are not competing purchases, and anyone who tells you otherwise is selling one of them.
Next
- Gage R&R calculator — paste a crossed study
- Cp/Cpk calculator — capability, once the gauge is trusted
- What is Cpk? — why %GRR has to come first