MSA and gauge R and R — AlpineTutor

Every number on a control chart, a capability report and a PPAP was produced by a measurement, and a measurement is a process with its own variation. Until you know how much of the spread belongs to the gauge, you do not know how much belongs to the part.

A metrology bench during a gauge study, with ten numbered parts arranged in a tray, a bore gauge and a digital micrometre on a stand, a data sheet recording three trials by three operators, and a set of gauge blocks with a calibration certificate beside them. — Ten numbered parts, three operators, three trials. The numbers on the parts are the part of the method people skip.

Three sources, three different fixes

A diagram decomposing observed total variation into its components. A wide band at the top labelled total observed variation splits into two bands, part to part variation and measurement system variation. The measurement system variation band splits further into repeatability, labelled equipment variation and illustrated by one operator measuring one part three times with slightly different readings, and reproducibility, labelled appraiser variation and illustrated by three operators each measuring the same part and obtaining three clustered but offset groups of readings. Beside repeatability a note reads worn or unsuitable equipment, and beside reproducibility a note reads undefined method, holding or reading differences. A separate box to the right shows bias as a systematic offset of all readings from a traceable reference value, and linearity as a bias that changes across the measuring range, with both marked as invisible within the study data alone. — The split down the middle is the study. The box on the right is what the study cannot see, and why calibration is a separate obligation.

Component	What it is	The fix
Repeatability	One operator, one part, repeated	The equipment
Reproducibility	Different operators, same parts	The method
Part-to-part	What is left	Nothing — this is the signal

High reproducibility with low repeatability is the most informative result you can get: each operator is consistent and they disagree with each other. That is holding, locating or reading — not the instrument. A better micrometre will change nothing; a fixture and a written method will fix it.

Reading the percentage

A diagram showing how a gauge study result is interpreted. A horizontal scale from zero to fifty per cent is divided into three bands, under ten per cent shaded as acceptable, ten to thirty per cent shaded as conditionally acceptable with a note listing the conditions of characteristic importance, cost of improvement and alternatives available, and over thirty per cent shaded as unacceptable. Below the scale, the same physical gauge spread of five micrometres is shown measured against three different tolerances, a wide tolerance of five hundred micrometres giving a small percentage, a moderate tolerance of fifty micrometres giving a mid band percentage, and a tight tolerance of twenty micrometres giving an unacceptable percentage, illustrating that the same gauge passes or fails depending on the characteristic. At the right, remedies are listed for an unacceptable result, averaging repeated readings, defining the method and fixturing, and guard banding the acceptance limits inward. — One gauge, three tolerances, three verdicts. The study belongs to the characteristic, not to the instrument.

Result	Verdict
Under 10 %	Acceptable
10 – 30 %	Conditional — on importance, cost, and what else exists
Over 30 %	Unacceptable; the measurement is dominating what you see

The middle band is conditional, not a pass. And the reason the result is a percentage is that adequacy is relative: a gauge with a five micrometre spread is excellent against a half millimetre tolerance and useless against twenty micrometres. The same instrument genuinely passes on one characteristic and fails on the next.

What the study cannot see

An R and R study measures spread. It is blind to a systematic offset, because a bias shifts every reading in the same direction and the numbers stay beautifully consistent.

Property	Question	Needs
Bias	Is the average offset from truth?	A traceable reference
Linearity	Does that offset change across the range?	References at several values
Stability	Does it change over time?	Repeated checks on a master

Linearity is the one that catches people out. A gauge can be accurate in the middle of its range and offset at the extremes — which is exactly where the parts you care about are. Checking bias at one convenient value near the middle proves nothing about the ends.

Attribute inspection: the kappa study

When the result is pass or fail rather than a number, there is no spread to analyse. Agreement is assessed instead: how consistently do inspectors agree with each other and with a known standard, beyond what chance alone would produce?

Visual inspection studied this way routinely returns agreement far below what anyone in the room expected, which is itself the argument for the prevention-based approach in error proofing and traceability.

Why this comes first

When there is no better gauge

Sometimes the result is unacceptable and nothing better exists at any sensible cost. There are legitimate responses:

Average repeated readings — this reduces the random component by a known factor.
Define the method — a fixture, a stated contact point, a written instruction. This attacks reproducibility directly.
Guard band the acceptance limits inward per ISO 14253-1 — deliberately accepting that good parts will be rejected in order to stop bad parts being accepted. The same logic as checking fixtures and functional gauges.

Recording the limitation and carrying on is not on that list.

What to take away

A measurement is a process with variation; the chart shows process plus gauge.
Repeatability is the equipment; reproducibility is the method.
Consistent operators who disagree with each other means fixture and method, not a new gauge.
The result is a percentage because adequacy is relative to the tolerance.
One gauge can pass on one characteristic and fail on the next.
Under 10 % acceptable, 10–30 % conditional, over 30 % unacceptable.
Part selection sets the denominator — span the real process range, sampled over time.
An R and R study is blind to bias; that needs a traceable reference.
Check linearity across the range, not at one comfortable value.
Attribute inspection needs a kappa study, and usually disappoints.
No better gauge? Average, fixture, guard band — and record the decision.