Sign in Start now

Measurement system analysis (MSA)

Last reviewed 2026-08-25

Measurement system analysis is how you find out how much of the variation in your data belongs to the measuring system rather than to the thing being measured. Until you know that number, every other number is provisional — a control chart, a capability index, a decision to scrap a batch.

The uncomfortable arithmetic is this. What you observe is not the process. It is the process plus the measurement system, and variances add:

σ²observed = σ²process + σ²measurement

A gauge contributing 30% of the observed standard deviation is contributing about 9% of the variance — which sounds tolerable until you notice that the capability number computed from the observed spread is now reporting a process worse than the one you have. The scrap that follows is real, and the process that caused it is not.

The five properties MSA measures

Most people meet MSA as "Gage R&R" and stop there. Gage R&R is one part of it. The full set:

Property The question it answers
Repeatability Same operator, same part, same gauge, twice — how close?
Reproducibility Different operators, same part — how close?
Bias Does the gauge read consistently high or low against a reference?
Linearity Does that bias change across the measurement range?
Stability Does any of the above drift over weeks?

Gage R&R covers the first two. It says nothing whatever about bias: a gauge reading 0.4 mm high on every part, every time, by every operator, produces an excellent Gage R&R result. Repeatable, reproducible, and wrong. Bias needs a reference standard, and no amount of R&R study substitutes for one.

Stability is the one most often skipped, because it needs the study repeated rather than run. A gauge that passed in March is evidence about March.

Gage R&R, and how to read it

The crossed ANOVA study — several operators, several parts, several repeats — partitions the observed variation into part-to-part, repeatability and reproducibility, including the operator × part interaction that the older average-and-range method cannot see and that people routinely drop from the Minitab dialog by accident. The full procedure, menu by menu, is here, and the free Gage R&R calculator runs the same analysis in a browser.

Two numbers come out, and they answer different questions.

%Study variation — gauge variation as a share of total study variation. The conventional reading, from the AIAG manual: under 10% acceptable, 10–30% conditional, over 30% unacceptable.

Number of distinct categories (ndc) — how many separate groups of parts the system can actually tell apart. Five or more is the usual threshold. An ndc of 2 means the gauge can sort parts into "big" and "small" and nothing finer, and a control chart drawn from it is a chart of the gauge.

The trap in %study variation

It is a ratio, and the denominator is the parts you chose for the study.

Pick ten parts that happen to be nearly identical and the part-to-part variation is small, so the gauge's share of the total is large, and a perfectly adequate gauge fails the study. Pick ten parts spanning far more than the process actually produces and the same gauge passes comfortably.

Neither result is about the gauge. The parts in an MSA study must span the real process range, and a study that does not say how its parts were selected cannot be interpreted at all. When the specification is what matters rather than the process spread, %tolerance is the more honest denominator — and it is a different number with a different meaning, so do not let a report quote one and label it the other.

Where MSA sits relative to everything else

The order matters and is frequently run backwards:

  1. Measurement system first. If the gauge cannot see the variation, nothing downstream is meaningful.
  2. Then stability. A control chart tells you whether the process is behaving. Limits computed on an unstable process describe nothing repeatable.
  3. Then capability. Cpk and Ppk assume a stable process measured by an adequate system. Computed before steps 1 and 2, a capability index is a number about nothing — and it will still look authoritative on the report.

A capability study run on an unvalidated gauge is the most common way a plant ends up confidently wrong about its own process.

Two related studies people reach for

Method comparison. When the question is whether two measuring methods agree rather than whether one is repeatable, a Gage R&R is the wrong tool — use a Bland-Altman plot, which shows bias and limits of agreement across the measurement range instead of a single summary statistic.

Correlation is not agreement. Two methods can correlate at r = 0.99 and disagree by a constant 4 units at every point. A scatter plot with the correlation is worth drawing, but never as evidence that two methods can be substituted for one another — that is precisely what Bland-Altman exists to test.

What the standards ask for

  • Automotive — the AIAG & VDA SPC manual and MSA reference manual sit behind IATF 16949, and an MSA study is expected as part of the PPAP package. See the standards page.
  • Aerospace — AS9103 asks for variation of key characteristics to be managed, which presupposes a measurement system capable of seeing it. More here.
  • Laboratory — the equivalent discipline is method validation and daily QC, where the control material's assigned mean and SD play the role the reference standard plays in industry. See Westgard rules and the Levey-Jennings chart.

The short version

Run the study on parts that span the real process range. Read ndc alongside %study variation, not instead of it. Remember that Gage R&R is silent on bias. Repeat it occasionally, because stability is a property of time and a study is a snapshot. And do it before the capability number, not after somebody has questioned one.

Run these on your own numbers

Free, no signup, and nothing you paste is stored — the same tested engine that draws the charts in the product, so the answers cannot disagree.

  • Gage R&R calculator (ANOVA) — Paste a crossed study — part, operator, reading — and get the full ANOVA: repeatability and reproducibility separated, the part-by-operator interaction tested rather than assumed away, %GRR, %Tolerance and ndc against AIAG's bands.
  • Bland-Altman plot — Paired readings from two methods, plotted as difference against average — with the bias, the limits of agreement, their confidence intervals, and a test for whether the disagreement depends on the magnitude.
  • Scatter diagram and correlation — Paired readings plotted, with Pearson r, Spearman rho, r², the least-squares line and its p-value — and the two reported together, because when they disagree the disagreement is the finding.
Open the calculator Start now

Read next

  • What is Cpk? — Cpk is the distance from your process mean to the nearer specification limit, measured in units of three sigma. Here is what it means, what a good value is, and the three ways it lies to you.
  • Cpk vs Ppk — what the gap between them is telling you — Cpk uses within-subgroup sigma, Ppk uses the standard deviation of every reading. When they disagree, the disagreement is the finding — not a rounding artefact to be argued away.
  • What is an SPC chart? — An SPC chart is a time-ordered plot with limits calculated from the process itself — never from the specification. That one distinction is what makes it a control chart rather than a line graph with lines on it.