Accuracy and precision: what can go wrong with a reading?

Every reading = the part's true value + the measurement error. MSA splits that error into two families.

Location = accuracy (trueness)

Is the average reading right?

Bias (at one point), linearity (bias across the range), stability (bias over time). Checked against a traceable reference value.

Width = precision

Do repeated readings agree?

Repeatability (same appraiser, same part) and reproducibility (between appraisers). This is what a gauge R&R study measures.

One part measured 40 times by one appraiser, with your current shop-floor settings

Question	Family	Study
Is the average reading right at one size?	Location
Is it right across the whole range of sizes?	Location
Does it stay right over weeks and months?	Location
Do repeated readings agree, within and between appraisers?	Width
The test destroys the part. How do I study the gauge?	Width
Go/no-go gauge or visual check: do inspectors make the same call?	Attribute
What does all this do to accept / reject decisions?	Result

A note on the words

The AIAG MSA manual uses "accuracy" for the location side. The international metrology vocabulary (VIM) and ISO 5725 call that part "trueness" and use "accuracy" for trueness and precision together. Both describe the same thing; we say "accuracy (trueness)" once and "accuracy" after that.

Before the study: will the result mean anything?

Four checks that decide whether any MSA study on this gauge can be trusted.

3. Where does the "true" value come from?

Bias and linearity are judged against a reference value. That value is only as good as its traceability chain:

National standardNPL India

→

Accredited labISO/IEC 17025, NABL

→

Your referencetoolroom gauge, CMM, slip gauges

→

Master partmeasured value

→

Gaugeunder study

Every link: valid certificate, stated uncertainty, due date
The parameter is inside the lab's accredited scope
The reference is clearly more accurate than the gauge under study

See ISO 9001 7.1.5.2 and IATF 16949 7.1.5.2.1, 7.1.5.3.1, 7.1.5.3.2.

Gauge R&R: how much does the gauge scatter?

Width only. The study estimates repeatability and reproducibility from the readings and compares them with the part variation and the tolerance.

Try a shop-floor problem

Where each appraiser's readings land

The charts, one at a time

Range chart by appraiser: is each appraiser consistent?

Look for: points above the red line. That appraiser did not get the same reading twice on that part.

Average chart by appraiser: can the gauge tell the parts apart?

Look for: points outside the red lines. Here that is good: the parts stand out from the gauge scatter.

Interaction plot: do appraisers rank the parts the same way?

Look for: lines that cross. Appraisers measure some parts differently.

All readings by part

Look for: one appraiser's dots sitting apart from the others on the same part.

How was this calculated? (with all the readings)

The readings

Quick check first? The range method

Bias: is the average reading right?

One master part with a known reference value, measured again and again. Bias = average reading − reference value. A t-test asks: could this bias be just scatter?

Try a shop-floor problem

The readings of the master part, with the reference value and the 95% confidence interval of the average

Look for: the reference line inside the 95% interval of the average means no real bias. Outside it, the bias is real.

Calibration certificate vs bias study

Is this master value trustworthy?

"The master's value is the drawing nominal."

Not traceable. The nominal is the target, not a measurement. Measure the master with a more accurate, calibrated instrument and record that value with its uncertainty.

"The master was measured once, three years ago."

Out of date. Masters wear, corrode and get dropped. Give the master its own calibration interval, like any gauge.

"The master was measured on the same type of micrometer we are studying."

Not good enough. The reference must be clearly more accurate than the gauge under study (many plants use 10 : 1, at least 4 : 1).

"The certificate is from an accredited lab, but its scope lists only pressure and temperature."

Not covered. Accreditation is per parameter and range. The lab's scope must include this kind of dimensional measurement.

"The certificate is valid, but its serial number does not match the master's ID."

No evidence for this master. The certificate belongs to another item, or the ID was changed. Same problem if the due date has passed.

How was this calculated?

Linearity: is the gauge right across its whole range?

Bias measured on 5 reference parts from small to large. A gauge can be right in the middle and wrong at the ends.

Try a shop-floor problem

Bias at each reference value, with the fitted line and its 95% confidence bands

Look for: the bias = 0 line staying inside the green band all the way across. Where it leaves the band, the gauge reads wrong at that size.

Calibration certificate vs linearity study

How was this calculated?

Stability: does the gauge stay right over time?

The same master part, checked a few times every day and plotted on an X̄ and R chart. Drift moves the average; a gauge getting worse widens the range.

Try a shop-floor problem

X̄ chart: is the average reading of the master moving?

Look for: points beyond the red lines, or a long run on one side: the gauge zero is moving.

R chart: is the scatter changing?

Look for: ranges growing, or beyond the red line: the gauge is getting worse.

Check average / rangeSignalReference value of the master

Destructive test: what if you cannot measure the same part twice?

Pull-off, peel, cross-section, hardness: the test uses up the piece. So "the same part" becomes "another piece from the same batch", and any difference between those pieces is counted as gauge scatter.

Try a shop-floor problem

What the study calls repeatability

All readings by batch: each dot is a different piece

Look for: how far apart the dots of one batch sit. All of that spread is counted as gauge scatter, even when it is the material.

Range chart: how much do pieces from one batch differ?

Look for: ranges above the red line: pieces in that batch differ more than in the others.

Which design? Shared batches or own batches

Batch shared by all (crossed)

Each batch is split: every appraiser tests some pieces from every batch. The batch plays the part's role, and the study is analysed like a normal gauge R&R.

Needs batches big enough for appraisers × pieces
Separates appraisers well
Can show a batch × appraiser interaction

Own batches per appraiser (nested)

Each appraiser tests their own batches. Used when a batch is too small to share, or appraisers work on different shifts or sites.

Any batch size works
Appraiser differences are mixed with which batches each one got
No interaction term

How was this calculated? (with all the readings)

The readings

Attribute agreement: do inspectors make the same accept / reject call?

Go/no-go gauges and visual checks give a call, not a number. The study checks whether each inspector agrees with themselves, with the others, and with the right answer.

Try a shop-floor problem

Every call, parts sorted from smallest to largest

Look for: wrong calls. Right next to a limit they are expected; far from the limits they point to the inspector or the gauge.

Near the lower limit: how often is each part accepted?

Look for: marks between 0% and 100%: the same part got different calls. One inspector's marks sitting apart from the others is a strict or lenient inspector.

Near the upper limit: how often is each part accepted?

Look for: marks between 0% and 100%: the same part got different calls. One inspector's marks sitting apart from the others is a strict or lenient inspector.

How was this calculated? (with all the calls)

The calls

What the gauge does to your decisions

What you measure = the parts' spread + the gauge's spread (+ its bias). Near the limits, the gauge decides whether a part passes.

Which gauge makes the call?

All parts: as they really are, and as the gauge shows them

Look for: the slivers at the limits: good parts rejected and bad parts accepted.

Zoom on the lower limit

Look for: amber = good parts rejected, red = bad parts accepted. A guard band moves the acceptance line inwards.

Zoom on the upper limit

Look for: amber = good parts rejected, red = bad parts accepted. A guard band moves the acceptance line inwards.

Try a guard band: accept only well inside the limits

Want your team to run these studies on your own gauges and parts, and act on the results? See how we run MSA training or let's talk about your measurement questions.