GAGE R&R · MEASUREMENT SYSTEM ANALYSIS · ANOVA METHOD · CROSSED OR NESTED · %TOLERANCE · NDC · CONFIDENCE INTERVALS · RUNS IN YOUR BROWSER

Is this gauge good enough to tell your parts apart?

Several operators measured the same parts several times. Paste the readings and get a Gage R&R by the ANOVA method: how much of the variation is the gauge (repeatability), how much is who held it (reproducibility), and how much is the parts themselves, as %study variation, %contribution, %tolerance and the number of distinct categories, each with a confidence interval so you can see how much ten parts and three operators can tell you. Crossed studies, or nested ones for destructive tests.

Nothing you enter leaves this browser tab. A statistical estimate from the readings you give, not engineering certification or quality-system advice.

Your gauge study

The example: is a JavaScript benchmark a capable gauge?

A gauge is anything that turns a thing into a number, and a stopwatch is one. To give this page a real study rather than a made-up one, ten workloads (sorting copies of arrays of 400 to 4,000 random numbers) were each timed three times by three "operators", three timing harnesses: one call between two performance.now() readings (single), fifty calls in a loop divided by fifty (loop50), and the median of fifteen single calls timed with process.hrtime.bigint() (median15), in shuffled order on one Node 22 process (the script is eval/benchmark-study.mjs). With the interaction kept, gauge R&R is 22.5% of study variation with 6 distinct categories: marginal. Repeatability is 15.7% and reproducibility 16.1%, and nearly all of the reproducibility is one harness: single reads 101.5 µs on average against 83.4 and 85.0 for the other two, the cost of timing one short call. Without it, gauge R&R falls to 14.7% and the distinct categories rise to 9.

Your own run of the script gives different numbers: that is the point of a gauge study. The bundled file is the one measured for this page.

Method and citation

For a crossed study with p parts, o operators and n trials, the two-factor random-effects ANOVA gives the mean squares for parts, operators, their interaction and repeatability, and the variance components follow from the expected mean squares (Montgomery and Runger 1993): repeatability is the error mean square; the interaction is (MSPO − MSE)/n; operator is (MSO − MSPO)/(pn); part-to-part is (MSP − MSPO)/(on). A negative estimate is reported as zero. Reproducibility is operator plus interaction, and gauge R&R is repeatability plus reproducibility. The nested design replaces the interaction with parts within operators. Intervals: the exact chi-square interval for repeatability, and modified large-sample intervals for the rest (Graybill and Wang 1980; Ting et al. 1990; Burdick, Borror and Montgomery 2003). The number of distinct categories is 1.41 × part SD / gauge SD, rounded down, and the 10% and 30% bands are the AIAG MSA manual's guidelines (4th edition, 2010).

Balanced studies only. Every operator must measure every part (or, nested, the same number of parts) the same number of times; the page says which cells are short.

Random operators. The intervals treat the operators as a sample from everyone who might use the gauge. With two or three operators the reproducibility interval is wide, and the page shows it.

Not here (yet): the average-and-range method, attribute agreement, bias and linearity studies, one-sided tolerances and unbalanced designs.

Checked against: sums of squares from statsmodels 0.15.0 OLS ANOVA and quantiles from scipy 1.18.1 on eight generated studies, to 1e-9; a hand-worked 2 × 2 × 2 study; and a coverage simulation of the intervals (README, "Verification").

@article{montgomery1993gauge,
  author = {Montgomery, Douglas C. and Runger, George C.},
  title = {Gauge capability and designed experiments. Part I: basic methods},
  journal = {Quality Engineering}, volume = {6}, number = {1}, pages = {115--135}, year = {1993},
  doi = {10.1080/08982119308918710}}
@article{burdick2003review,
  author = {Burdick, Richard K. and Borror, Connie M. and Montgomery, Douglas C.},
  title = {A review of methods for measurement systems capability analysis},
  journal = {Journal of Quality Technology}, volume = {35}, number = {4}, pages = {342--354}, year = {2003},
  doi = {10.1080/00224065.2003.11980232}}
@article{ting1990mls,
  author = {Ting, Naitee and Burdick, Richard K. and Graybill, Franklin A. and Jeyaratnam, S. and Lu, T.-F. C.},
  title = {Confidence intervals on linear combinations of variance components that are unrestricted in sign},
  journal = {Journal of Statistical Computation and Simulation}, volume = {35}, number = {3-4}, pages = {135--143}, year = {1990},
  doi = {10.1080/00949659008811240}}
@article{graybill1980mls,
  author = {Graybill, Franklin A. and Wang, Chih-Ming},
  title = {Confidence intervals on nonnegative linear combinations of variances},
  journal = {Journal of the American Statistical Association}, volume = {75}, number = {372}, pages = {869--873}, year = {1980},
  doi = {10.1080/01621459.1980.10477565}}
@manual{aiag2010msa,
  title = {Measurement Systems Analysis Reference Manual}, edition = {4th},
  organization = {Automotive Industry Action Group}, address = {Southfield, Michigan}, year = {2010}}

Questions

What do %study variation, %contribution and %tolerance mean?

%contribution is each source's share of the total variance. %study variation is its standard deviation divided by the total standard deviation, so the column does not add to 100. %tolerance is the source's study variation (6 standard deviations by default) divided by the tolerance, USL minus LSL; for gauge R&R it is the precision-to-tolerance ratio. Use %tolerance when the gauge sorts parts against a specification, and %study variation (or %process with a historical process standard deviation) when it is used to study or control the process.

When is a gauge acceptable?

The AIAG Measurement Systems Analysis manual's guideline is under 10% gauge R&R acceptable, 10% to 30% possibly acceptable depending on the application and costs, and over 30% unacceptable, with at least 5 distinct categories. These are guidelines, not laws of nature: the intervals on this page show how far the study itself could be off, and with 10 parts and 3 operators that is often a lot.

What is the number of distinct categories?

It is 1.41 times the part-to-part standard deviation divided by the gauge R&R standard deviation, rounded down: roughly how many non-overlapping groups of parts the gauge can tell apart across the parts in the study. It depends on the parts you chose, so a study with parts that span less than the process range will understate it.

Crossed or nested?

Crossed when every operator measures every part several times. Nested when each part can be measured by only one operator, as in destructive testing (a tensile test, a burst test), where each operator gets their own parts from a batch assumed to be uniform. In a nested study any real difference between batches shows up as reproducibility.

What does pooling the interaction do?

In a crossed study the part by operator interaction measures whether operators disagree by different amounts on different parts. When its F test is not significant, a common practice is to drop it and pool its sum of squares into repeatability, which gives the operator and part tests more degrees of freedom. The default here pools when p is above 0.05; choose keep to always report it, or set your own threshold.

How are the confidence intervals computed?

Repeatability has an exact chi-square interval. The other components are linear combinations of mean squares with positive and negative coefficients, and get the modified large-sample (MLS) intervals of Graybill and Wang (1980) and Ting et al. (1990), as reviewed for gauge studies by Burdick, Borror and Montgomery (2003). In a simulation of 2,000 studies for each of seven designs, the 95% intervals covered the true value 94.5% to 97.7% of the time.

Is my data uploaded anywhere?

No. Every number is computed by JavaScript in this browser tab. There are no accounts, cookies or analytics vendors.