Gauge R&R Study

What this tool does

CM8-210 runs a measurement system analysis — a gauge R&R study — by the average and range method. Record one row per reading: which part, which appraiser, which trial, what the gauge said. The tool works out repeatability, reproducibility, the combined gauge R&R, the part variation and the total variation; expresses the gauge error as a percentage of the study variation and of your tolerance; counts the distinct categories the system can resolve; and prints a study you can put in front of an auditor.

Everything happens inside this single file — no account, no upload, no network request. The study is computed from every reading in the register rather than from whatever the filters show, because a gauge R&R calculated from part of a study is not a gauge R&R.

What MSA actually answers

Before you argue about whether the process is capable, find out whether you can measure it. Every reading is the true value plus measurement error, and you never see the two separately. A gauge consuming 30 % of your tolerance makes every other number suspect: capability indices computed from it are wrong, control charts react to the gauge rather than the process, and arguments about whether a batch conformed become arguments about noise.

The consequences run both ways. Good parts get rejected, bad parts get accepted, operators lose faith and start measuring twice, and improvement projects chase variation that is not in the process at all. A study takes a morning; the arguments it prevents take months.

Repeatability and reproducibility

The two halves answer different questions, and the whole diagnostic value of the study is in knowing which is larger.

Repeatability is the same appraiser, the same part, measured again — equipment variation, what the instrument and the method do when nothing else changes. When it dominates, look at the physical things: a worn or unstable gauge, a fixture that does not locate the part the same way twice, too little resolution on the display, a surface the gauge cannot read consistently.

Reproducibility is different appraisers, the same part — appraiser variation, the disagreement between the people. When it dominates, the instrument is probably fine and the problem is human: training, an ambiguous work instruction, no agreement on where to measure or how firmly to seat the part. That is usually the cheaper half to fix, and the tool names the dominant half on the tiles.

The average and range method

Each appraiser measures each part two or three times. For every part-and-appraiser cell the tool takes the range across the trials; the mean of those ranges is R-bar. Everything follows from it:

Two details matter. The subtraction inside AV removes the repeatability noise the appraiser means already contain, so the tool does not blame the people for the gauge; a negative result is floored at zero. And the components combine as squares: an EV of 40 % and an AV of 30 % give a GRR of 50 %, not 70 %. The R&R summary table prints every intermediate figure with its arithmetic beside it.

%study variation and %tolerance

The same gauge error reported two ways, answering different questions.

%GRR of study variation compares the measurement error with the spread of the parts. Use it when improving a process: it says whether the gauge can see the differences you are trying to reduce. Its weakness is the denominator — parts all taken from one hour of one shift barely differ, part variation collapses, and a good gauge scores badly. If you know the real six-sigma spread of the process, enter it in the settings and the percentages are measured against that instead.

%GRR of tolerance compares the measurement error with the specification band. Use it when inspecting to a spec: it says how much of your tolerance the gauge eats before the process has used any. It does not depend on which parts you picked, which makes it more robust — but it is silent about whether the gauge can support improvement.

The conventional thresholds, changeable in the settings: at or under 10 % acceptable; 10–30 % marginal, depending on the importance of the characteristic and the cost of being wrong; over 30 % not acceptable. Conventions, not laws.

Distinct categories

ndc asks a blunt question: how many separate groups can this system actually sort the parts into? An ndc of 2 means the gauge can tell you "big" or "small" and nothing more. Five or more is the usual bar, because five categories are the minimum needed to estimate a distribution, plot a meaningful control chart or calculate a capability index that means anything. ndc and %GRR are two views of one ratio, so they normally agree — but ndc is the easier one to explain.

Running the study properly

The arithmetic is trivial. Getting a truthful answer is entirely about how the readings were taken.

• Use the same physical parts throughout. Number them, and make sure part 3 is the same lump of metal for everybody. A study where each appraiser measured "some parts" measures nothing.
• Pick parts that span the real process range — not scrap, not hand-picked perfection, but the ordinary spread the process produces.
• Randomise the order within each round, and re-randomise between rounds.
• Keep the appraisers blind to their previous readings and to each other's; people who can see what they wrote last time write it again.
• Use the real appraisers, in the normal place, not the best inspector at a quiet bench.
• Calibrate first. R&R measures precision, not accuracy: a gauge reading 0.02 high on everything passes with flying colours.
• Record every digit the gauge resolves. Rounding is itself measurement error.

The standard design is ten parts, three appraisers, two trials. Fewer still produces a number — the tool carries constants down to five parts and two appraisers — but the estimate gets rougher, and the study design tile says so.

What this method cannot do

This is the average and range method, sometimes called the long form: the method most quality manuals describe, checkable with a calculator, honest arithmetic — but not the best available estimate.

The ANOVA method estimates the same quantities better, and does one thing this method cannot: it separates the part-by-appraiser interaction, the case where one appraiser struggles with one particular part rather than reading everything high or low. The average and range method folds any interaction into reproducibility, so it reports "the appraisers disagree" without telling you the disagreement is confined to two awkward parts. The cell-by-cell repeatability chart is the nearest this tool gets, and it is often enough to spot the pattern by eye. The constants also assume a balanced study. Fill in every cell.

What to do with the result

A study that passes lets you trust everything downstream. Feed the same characteristic into the Process Capability (Cp/Cpk) tool to find out whether the process can hold the tolerance, and into the SPC Control Chart Builder to watch it over time — both are only as meaningful as the measurement system underneath them.

A study that fails tells you where to spend. Repeatability dominant: fix the gauge, the fixture or the method, or buy resolution. Reproducibility dominant: write the work instruction properly, agree exactly how and where the measurement is taken, train to it — then re-run the study, the only proof the fix worked.

The spreadsheet workflow

If the readings already live in a spreadsheet — and R&R data usually does — there is no need to retype them.

• Spreadsheet template in the toolbar saves a CSV whose headings are exactly this tool's column names — study name, gauge or instrument, characteristic, part number, appraiser, trial, measurement, study date, notes — with a guidance row underneath showing what each column expects.
• Fill in one row per reading and delete the guidance row before saving as CSV. Most R&R sheets are a grid of parts against appraisers; unstack that grid into one reading per row.
• Import spreadsheet reads it back. Columns are matched by heading, so their order does not matter and extra columns are ignored. Rows missing a required column, or failing a validation check, are skipped and reported by row number.

The file is read by this page in your browser: nothing is uploaded and the original is not changed. Importing adds to what is here rather than replacing it.

FAQ

Can I run it with two parts? No. The part constant starts at five; below that the part variation rests on so little evidence that the percentage would be meaningless, so the tool says so rather than printing a confident wrong number.

My %GRR is terrible but the gauge is new. Check the parts before you blame the gauge. If they barely differ, part variation is small and %GRR of study variation is large by construction. Read %GRR of tolerance instead, or enter the known process variation.

Reproducibility came out as zero. The appraiser means agreed to within the gauge's own scatter, so the calculation would have gone negative and is floored at zero. The people are not the problem.

Does this replace calibration? No. Calibration is accuracy against a standard; R&R is precision. You need both, calibration first.

How often should I repeat it? When the gauge, method or people change, after a repair, and periodically for characteristics that matter. A study from three years ago describes a system that may no longer exist.

Saving your work

Readings, settings and the report header are written to this browser's local storage as you type. That storage belongs to one browser on one computer: another browser, a private window or a clean-up tool that clears site data will not have it.

Treat Export .json as the real save — one file containing everything, which Import .json restores anywhere. Export CSV gives you every filtered reading for spreadsheet work. Reset asks twice, then erases everything. There is no undo.

Accuracy & disclaimer

The arithmetic is elementary and the tool performs it faithfully. Everything that decides whether the answer is true sits underneath it: whether the same physical parts were used, whether they spanned the real process range, whether the order was randomised, whether appraisers could see their previous readings, whether the gauge was calibrated first.

A study with fewer than ten parts and three appraisers gives a rough estimate. The constants assume the standard balanced design and are approximations even then, and this method cannot separate the part-by-appraiser interaction. The result is indicative, not a certificate — a calculation aid, not a gauge calibration, and not a substitute for whatever method your customer or accreditation body requires.