When measuring destroys the part
A gauge study asks a simple question: when the number changes, is that the part changing or the measurement changing? The standard answer is to have several operators measure the same parts several times each, and split the variation. It works — right up until the measurement consumes the part.
The problem with a tensile test. Pull a specimen to failure and you cannot pull it again. Same for a hardness indent on a finished face, a chemical analysis that dissolves the sample, a burst test. There is no such thing as "the same part measured twice", so there is no such thing as the crossed study the textbook describes.
What people do instead, and why it goes wrong. The usual workaround is to give each operator their own set of parts and run the numbers as if it were an ordinary study. But then the software believes that operator A's part 1 and operator B's part 1 are the same part. They are not — they are two different pieces of steel. Every genuine difference between them gets charged to the measurement system. The %GRR comes out enormous and the gauge gets condemned for the crime of the parts being different.
How big is the error? On a simulated study SenSight reports 18.9% for the correct nested analysis and 92.5% for the same measurements analysed as crossed. Repeatability is identical in both — 0.382 — because repeats within a part really are repeats. The whole difference sits in the part variation: 8.58 when read correctly, 3.36 when the model is wrong. One reading says "usable with care"; the other says "replace the equipment".
The nested design. The honest structure is that parts sit inside operators: each appraiser gets their own specimens, cut adjacent to each other so they are as alike as you can make them, and measures each one two or three times. SenSight reads which structure your data has rather than asking you to declare it — if every part code appears under exactly one operator, it is nested, and that is how it is analysed.
What you give up. In a nested study the operator×part interaction cannot be estimated. Not "is assumed zero" — cannot be estimated, because separating it would require the same part measured by two people, which is exactly what you do not have. Any tool that prints an interaction term for destructive data is making it up.
And one thing that has nothing to do with destruction. Even for ordinary studies, ANOVA beats the average-range method, because the range method cannot separate that interaction either. If one operator reads the thick parts high and the thin parts low, the range method buries it in repeatability and you go off retraining people when the real fault is in how the part is located in the fixture.
The problem with a tensile test. Pull a specimen to failure and you cannot pull it again. Same for a hardness indent on a finished face, a chemical analysis that dissolves the sample, a burst test. There is no such thing as "the same part measured twice", so there is no such thing as the crossed study the textbook describes.
What people do instead, and why it goes wrong. The usual workaround is to give each operator their own set of parts and run the numbers as if it were an ordinary study. But then the software believes that operator A's part 1 and operator B's part 1 are the same part. They are not — they are two different pieces of steel. Every genuine difference between them gets charged to the measurement system. The %GRR comes out enormous and the gauge gets condemned for the crime of the parts being different.
How big is the error? On a simulated study SenSight reports 18.9% for the correct nested analysis and 92.5% for the same measurements analysed as crossed. Repeatability is identical in both — 0.382 — because repeats within a part really are repeats. The whole difference sits in the part variation: 8.58 when read correctly, 3.36 when the model is wrong. One reading says "usable with care"; the other says "replace the equipment".
The nested design. The honest structure is that parts sit inside operators: each appraiser gets their own specimens, cut adjacent to each other so they are as alike as you can make them, and measures each one two or three times. SenSight reads which structure your data has rather than asking you to declare it — if every part code appears under exactly one operator, it is nested, and that is how it is analysed.
What you give up. In a nested study the operator×part interaction cannot be estimated. Not "is assumed zero" — cannot be estimated, because separating it would require the same part measured by two people, which is exactly what you do not have. Any tool that prints an interaction term for destructive data is making it up.
And one thing that has nothing to do with destruction. Even for ordinary studies, ANOVA beats the average-range method, because the range method cannot separate that interaction either. If one operator reads the thick parts high and the thin parts low, the range method buries it in repeatability and you go off retraining people when the real fault is in how the part is located in the fixture.
Try it in the app
Try: Gage R&R tab, ANOVA panel. Pick your part and operator columns and read the design line before running — SenSight tells you whether your data is crossed or nested before it computes anything.