Measurement System Analysis (MSA) is a set of studies that check whether a gauge and the way it is used can be trusted before you trust the numbers it produces. It quantifies how much of the variation you see comes from the parts and how much comes from the measurement itself, across bias, linearity, stability, repeatability, and reproducibility.
Every capability index, control chart, and pass/fail decision rests on measured data, and measured data is never perfectly true. If the measurement system adds too much of its own noise, a good process can look bad, a bad process can pass, and no amount of downstream analysis will fix it. MSA is the discipline that proves the gauge is honest first. This post covers the five properties MSA checks, how a gage R&R study works, how to read the numbers, and why MSA is a different question than process capability.
What is measurement system analysis?
Measurement system analysis is the study of the measurement process itself, the gauge, the operator, the method, the fixture, and the environment, to separate real part-to-part variation from variation the measurement adds. The core idea is that every reading is the true value plus measurement error, and MSA estimates how large that error is relative to the tolerance and to the process spread.
A measurement system has two families of error. Accuracy problems shift readings away from the true value: bias (a consistent offset), linearity (bias that changes across the range), and stability (bias that drifts over time). Precision problems scatter readings around a value even when nothing changed: repeatability (same operator, same part, same gauge) and reproducibility (different operators or setups on the same part). A trustworthy gauge has to be good on both.
By the numbers. The Automotive Industry Action Group's MSA reference manual sets the widely used acceptance rule for gage repeatability and reproducibility: a measurement system with a %GRR under 10% of the study variation or tolerance is generally acceptable, 10–30% may be acceptable depending on the application, cost, and risk, and over 30% is unacceptable and needs work (AIAG, Measurement Systems Analysis). The same manual asks for a number of distinct categories (ndc) of at least 5, meaning the system can reliably tell apart at least five groups of parts across the process range (ASQ, Measurement Systems Analysis).
What are the five properties MSA checks?
MSA breaks measurement quality into five properties, three about location and two about spread. You do not always study all five for every gauge, but a full MSA on a critical characteristic covers each.
- Bias the difference between the average of many measurements of a known standard and the standard's true value. A caliper that reads 0.02 mm high on a certified gauge block has a bias.
- Linearity whether bias stays constant across the measurement range. A gauge accurate at the low end but off at the high end has a linearity problem.
- Stability whether bias holds over time. You track it by measuring the same reference part periodically and plotting the result on a control chart.
- Repeatability the variation when one operator measures the same part many times with the same gauge. This is the gauge's own inherent scatter, sometimes called equipment variation.
- Reproducibility the variation added when different operators, or different setups, measure the same parts. This is usually a method or training problem, sometimes called appraiser variation.
What is gage R&R?
Gage R&R (gage repeatability and reproducibility) is the most common MSA study, and it targets the two precision properties. It answers one question: of all the variation you observe when measuring parts, how much is real part-to-part difference and how much is the measurement system's own noise? The study splits total variation into part variation and measurement variation, and measurement variation into repeatability plus reproducibility.
The reason gage R&R gets the most attention is that precision error is what quietly corrupts everyday decisions. If the gauge's spread is a big share of the tolerance, parts near a spec limit get sorted almost at random, good ones rejected, bad ones shipped. A capability study built on such data is meaningless, which is why gage R&R is the first check in any serious study of process capability (Cp and Cpk).
A quick example shows the stakes. Suppose a bore has a tolerance of 0.10 mm and the gauge's measurement spread is 0.04 mm. That measurement noise eats 40% of the tolerance band before you have made a single part, so any reading within 0.02 mm of a limit is effectively a coin flip. Cut the measurement spread to 0.01 mm and the same limit is now clean; the identical process suddenly looks far more capable, because the gauge stopped adding fog to the picture.
How do you run a gage R&R study?
A standard crossed gage R&R has every operator measure every part several times, in a randomized order. The classic recipe is 10 parts, 3 operators, and 2 or 3 trials each, but the principle matters more than the exact counts.
- Pick parts that span the process range. Choose about 10 parts that cover the normal spread of production, not 10 near-identical good ones. If the parts barely differ, the study cannot see part variation and every ratio looks bad.
- Calibrate and confirm the gauge first. Verify calibration and check bias against a known standard. Gage R&R measures precision; it assumes accuracy has already been handled separately.
- Have each operator measure blind and in random order. Two or three appraisers each measure all parts, two or three times, without seeing prior readings or part labels. Randomizing order keeps memory and drift from biasing the trials.
- Compute the variance components. Use ANOVA (or the average-and-range method) to split total variation into part-to-part, repeatability, and reproducibility, then express measurement variation as %GRR of the tolerance or of total study variation.
- Judge, fix, and repeat if needed. Compare %GRR and ndc to the acceptance rules. If the system fails, attack the larger component, repeatability points at the gauge or fixture, reproducibility points at method and training, then rerun.
How do you read the results?
Two numbers carry most of the verdict: %GRR and the number of distinct categories. %GRR is measurement variation as a percentage of tolerance or total variation, and lower is better. The ndc estimates how many separate groups of parts the system can reliably distinguish across the range; five or more means the gauge has enough resolution to be useful for process control.
| %GRR | Verdict (AIAG guideline) | number of distinct categories |
|---|---|---|
| Under 10% | Acceptable | 5 or more, acceptable |
| 10% to 30% | May be acceptable by application, cost, and risk | Watch resolution |
| Over 30% | Unacceptable, improve the system | Under 5, inadequate |
A trap worth naming: %GRR against the tolerance and %GRR against total study variation answer different questions. The tolerance basis asks whether the gauge is good enough to sort to spec; the total-variation basis asks whether it is good enough to study the process. A gauge can pass one and fail the other, so state which basis you used.
Resolution deserves its own check before you even run the study. A gauge should be able to read to at least one-tenth of the process spread or the tolerance, whichever is tighter; this is sometimes called the ten-to-one or discrimination rule. If a caliper reads only to 0.01 mm on a feature that varies by 0.03 mm, it can distinguish barely three steps, and no amount of averaging will rescue an ndc that low. Fix resolution first, because a gauge that cannot see the variation will fail gage R&R no matter how steady the operators are.
How is MSA different from process capability?
MSA and a capability study look at the same data stream but ask opposite questions. A capability study assumes the measurements are trustworthy and asks whether the process fits inside the spec. MSA questions that assumption and asks whether the measurement system is trustworthy in the first place. You cannot skip MSA and go straight to Cpk, because a Cpk computed on noisy measurements is a number about the gauge as much as the process.
The order is fixed: prove the measurement system with MSA, then establish that the process is stable on a control chart and part of a working statistical process control program, then compute capability. Skip the first step and every number after it inherits the gauge's noise. This sequencing is a small but non-negotiable piece of building real quality into a plant, a foundation under lean and continuous improvement, not a paperwork exercise. When calibration status, operator, and reading history are captured live on the floor rather than reconstructed from clipboards, MSA stops being an annual scramble and becomes a standing check, which is the kind of connected quality data Harmony surfaces in the CLS case study. For the specialized deep dive, see gage R&R.