Quality Engineering · Measurement System Analysis · Beginner
Gage R&R
You Can See
A measurement is never just a number. Learn to see what came from the part, what came from the gage, and what came from the person using it—then run the study yourself.
What should create the spread?
A good system makes real part differences easy to see and measurement noise hard to notice.
Start With the Question, Not the Formula
Vocabulary note: industry software and the AIAG manual commonly spell the word gage. General English often uses gauge. This lesson uses “Gage R&R,” but both spellings refer to the same study.
Every Reading Has a Story
Suppose a pipe wall is truly 8.000 mm thick. A micrometer might show 8.006 mm. The extra 0.006 mm did not necessarily come from the pipe.
Observed reading = reference or part value + systematic error + random errorMeasurement Anatomy Lab
Keep the same part. Change only the measurement system.
Systematic Error and Random Error Are Different
Shift or pattern
The results are consistently high, low, changing with size, or drifting with time. Bias, linearity, and stability problems live here.
It can be precise and still wrong.
Scatter
The result moves unpredictably when the same thing is measured again. Repeatability and reproducibility are major sources of this scatter.
It makes real differences harder to see.
Accuracy and Precision Are Not Synonyms
Gage R&R mainly studies precision. Bias and linearity studies assess accuracy against reference values. A complete MSA needs both questions.
Six Concepts Beginners Must Separate
Select a card. Each concept asks a different question and needs different evidence.
Repeatability (equipment variation, EV)
Question: If one appraiser measures the same part several times under the same conditions, how tightly do the readings agree?
Evidence: repeated readings, within-cell ranges, the R chart, and the repeatability variance component.
Typical action: inspect gage condition, fixturing, locating method, measurement force, environment, and resolution.
Where Part Variation Fits
Repeatability
Scatter when the same appraiser repeats the measurement.
Reproducibility
Differences among appraisers, plus part × appraiser interaction when included.
Total Gage R&R
The combined measurement-system standard deviation.
Do not add standard deviations directly. Independent variance components add. Convert back to standard deviation by taking the square root.
Bias, Linearity, and Stability—Made Visible
Bias and Linearity Lab
Bias is the average measurement minus the accepted reference value. Positive bias means the gage reads high; negative bias means it reads low. Linearity asks whether that bias stays the same across the operating range.
Reference-to-Reading Simulator
A healthy fitted line is close to horizontal and close to zero.
How the formal study works: choose reference parts across the gage range, obtain accepted reference values, have one appraiser measure each part repeatedly in randomized order, and regress bias on reference value. Use the slope test for linearity and the bias tests—not a visual threshold alone.
Stability Lab
Stability means the measurement system stays statistically consistent over time. A gage can pass a short Gage R&R today and drift next month.
Check-Standard Timeline
The same reference is measured three times each period.
| Question | Concept | Typical study |
|---|---|---|
| Do repeated readings agree? | Repeatability | Type 1 or Gage R&R |
| Do appraisers agree? | Reproducibility | Crossed/nested Gage R&R |
| Is the average correct? | Bias | Type 1 or bias study |
| Is accuracy consistent across size? | Linearity | Gage linearity and bias study |
| Does performance stay consistent over time? | Stability | Check standard + control chart |
Study Design Comes Before Software
Pick a situation
Crossed Gage R&R
Every appraiser measures every part. Use this for non-destructive measurements such as wall thickness on the same pipe rings.
Structure: Part × Appraiser × Trial
A Defensible Crossed Study—Step by Step
- Define the decision. State the characteristic, units, specification, intended use, and required measurement performance.
- Select representative parts. Use at least about 10 parts spanning the normal low-to-high process range—not ten nearly identical parts.
- Select typical appraisers. A common starting point is three people representing normal shifts, skill levels, and methods.
- Plan repeats. Use two or three trials per appraiser-part cell. A 10 × 3 × 2 design gives 60 readings.
- Randomize and blind. Hide part identity where practical, randomize order for every trial, and keep earlier results hidden.
- Reset the measurement. Remove, reposition, and re-fixture the part between repeats. Re-reading the same display understates error.
- Use normal conditions. Keep the production method, fixture, environment, instrument, and rounding rule realistic.
- Record—not erase—events. Document unusual conditions. Investigate questionable readings rather than silently deleting them.
Destructive-test warning: if appraisers do not measure the same physical parts, the ordinary crossed model is not present. Use a justified nested or expanded design; do not relabel different specimens as the “same part.”
Run a Gage R&R Study in Real Time
This miniature measurement lab generates a balanced crossed study. Change one source of variation and watch every graph and metric respond.
Pipe-Wall Measurement Control Room
Nominal wall = 8.000 mm. Each simulated reading contains part, appraiser, interaction, and repeatability effects.
Now change one error source or play the randomized collection to see what the analysis can and cannot know before all planned readings exist.
Measurements by part
Healthy: tight clusters within each part and clear separation between part averages.
Part × appraiser interaction
Healthy: appraiser lines are close and roughly parallel.
Variance contribution
Healthy: real part-to-part variance dominates the measurement components.
Appraiser comparison
Healthy: appraiser centres and spreads are similar.
Within-cell range chart
Each point is the range of all repeated readings for one completed part-appraiser cell. A point appears only after every planned trial in that cell has been collected.
Open the randomized measurement table
| Order | Part | Appraiser | Trial | Measurement (mm) | Status |
|---|
Which graph exposes poor repeatability?
Load the poor-repeatability preset. Predict what changes before showing the answer.
What does crossing mean?
Load interaction problem. Why can similar overall appraiser averages still mislead?
Why can narrow parts hurt ndc?
Load narrow part range. The gage did not change—so why did the study look worse?
Four Metrics, Four Different Questions
| Metric | Question answered | Common guideline | Important caution |
|---|---|---|---|
| %Study Variation | How large is measurement spread relative to the spread in this study? | <10% acceptable; 10–30% conditional; >30% generally unacceptable | Depends strongly on whether study parts represent the intended process variation. |
| %Tolerance | How much of USL − LSL is consumed by 6σ of measurement error? | Same common 10% and 30% bands | Needs the correct full tolerance and the chosen study-variation multiplier. |
| %Contribution | What share of total variance comes from each source? | <1% acceptable; 1–9% conditional; >9% generally unacceptable for Total Gage R&R | Uses variance, so it is the squared counterpart of the study-variation proportion. |
| ndc | How many distinct part categories can the system separate? | Common minimum ≥5; ≥10 is stronger | A narrow part sample can reduce ndc even when repeatability is unchanged. |
These are guidelines, not universal laws. Customer rules, decision risk, product safety, cost, and intended use can require stricter criteria. State the governing criterion in the report.
Core Formulas
%Study Variation
\(100\times\sigma_{GRR}/\sigma_{total}\)Compares measurement spread with the observed study spread.
%Tolerance
\(100\times 6\sigma_{GRR}/(USL-LSL)\)Uses the full tolerance width. Minitab’s default study-variation multiplier is 6.
%Contribution
\(100\times\sigma^2_{GRR}/\sigma^2_{total}\)Compares variance rather than standard deviation.
Distinct categories
\(ndc=1.41\times\sigma_{part}/\sigma_{GRR}\)Software reports an integer category count.
Result Interpreter
Enter standard-deviation components from an analysis.
Enter valid non-negative components and a positive tolerance.
Read the Graphs Before the Verdict
- Start with the R chart or within-cell ranges. Unstable or inconsistent repeats make later averages less trustworthy.
- Check measurements by part. Repeats should cluster tightly; part averages should cover the intended range.
- Check measurements by appraiser. Look for shifted centres, unequal spread, rounding patterns, or one appraiser behaving differently.
- Check part × appraiser interaction. Crossing or fanning lines mean disagreement depends on the part.
- Then read variance components and metrics. Identify the dominant measurement source and connect it to action.
Pipe-Wall Micrometer Study
A quality team studies a 0.500 mm-wide tolerance using 10 pipe rings, 3 inspectors, and 2 randomized trials. ANOVA estimates:
Run It in Minitab or Build Carefully in Excel
Minitab · Crossed
Stat → Quality Tools → Gage Study → Gage R&R Study (Crossed)- Place one reading per row in Part, Appraiser, and Measurement columns.
- Assign the matching columns to Part numbers, Operators, and Measurement data.
- Select ANOVA as the method.
- In Options, enter USL − LSL as process tolerance and confirm the study-variation multiplier.
- Request confidence intervals when decisions are high risk.
Minitab · Other designs
Stat → Quality Tools → Gage Study- Gage R&R Study (Nested): parts are unique to appraisers.
- Gage R&R Study (Expanded): additional factors, fixed factors, crossed/nested mixtures, or unbalanced data.
- Type 1 Gage Study: one reference part, one appraiser, repeated measurements.
- Gage Linearity and Bias Study: accuracy across reference values.
Recommended Data Layout
| Part | Appraiser | Trial | Measurement_mm | Random_Order |
|---|---|---|---|---|
| P07 | A | 1 | 8.214 | 1 |
| P02 | A | 1 | 7.986 | 2 |
| P09 | C | 2 | 8.331 | 37 |
Never average the trials before analysis. The repeats are needed to estimate repeatability.
Excel Function Table
| Task | Excel function or formula | Use | Caution |
|---|---|---|---|
| Cell average | AVERAGEIFS() | Average by part and appraiser | Needs tidy identifiers and does not estimate variance components by itself. |
| Within-cell spread | MAXIFS()-MINIFS() | Range of repeated measurements | Range method is less flexible than ANOVA, especially with interaction. |
| Sample standard deviation | STDEV.S() | Descriptive scatter | Do not substitute one overall SD for a crossed variance-component model. |
| Root-sum-of-squares | SQRT(EV^2+AV^2+INT^2) | Combine standard-deviation components | Only after valid components have been estimated. |
| %Study Variation | 100*GRR/SQRT(GRR^2+PV^2) | Relative to study variation | PV depends on representative part sampling. |
| %Tolerance | 100*6*GRR/(USL-LSL) | Relative to specification width | Confirm the multiplier and full tolerance. |
| ndc | MAX(1,INT(1.41*PV/GRR)) | Distinct category estimate | Handle GRR = 0 separately. |
| Bias | AVERAGE(Measurements)-Reference | Average offset | Formal inference needs repeated reference measurements. |
| Linearity slope | SLOPE(BiasRange,ReferenceRange) | Bias change across range | Use the formal study output for tests and confidence intervals. |
Excel limitation: ordinary worksheet formulas do not automatically fit the required random-effects ANOVA or handle pooling, interaction, negative variance estimates, unbalanced data, or confidence intervals correctly. Use a validated workbook or statistical software for formal decisions. The Average-and-Range method is useful for teaching and simple balanced studies, but ANOVA is stronger when interaction matters.
Match the Symptom to the Correct Action
| Evidence | Likely source | First investigation |
|---|---|---|
| Large within-part scatter; high ranges | Repeatability | Gage wear, dirt, contact force, fixture, location, environment, resolution. |
| One appraiser is consistently higher | Reproducibility | Method interpretation, zeroing, pressure, training, setup, fixture use. |
| Appraiser lines cross or fan | Part × appraiser interaction | Part-specific access, surface shape, technique, datum selection, readability. |
| All averages shifted from references | Bias | Calibration, zero offset, master value, environmental correction. |
| Bias becomes larger at high values | Linearity | Calibration range, scale factor, worn mechanism, wrong correction model. |
| Check standard trends over weeks | Stability | Wear, temperature, maintenance, master drift, recalibration interval. |
| Few distinct displayed values | Resolution | Instrument discrimination or excessive rounding/data entry. |
Six Traps That Ruin Conclusions
Narrow part range
Understates PV, inflates %Study Variation, and reduces ndc.
Wrong design
Crossed analysis cannot represent destructive measurements on unique specimens.
No repositioning
Repeated display readings make repeatability look artificially small.
Adding SDs
Variance components add; standard deviations combine by root-sum-of-squares.
Ignoring graphs
A single Total Gage R&R value can hide interaction or an abnormal appraiser.
Equating precision with accuracy
A highly repeatable gage can still be biased or nonlinear.
Practice case: Should this system be released?
A study reports %Study Variation = 8%, %Tolerance = 7%, and ndc = 12. A separate reference study shows +0.080 mm bias on a ±0.100 mm specification.
Answer: Do not release it based on Gage R&R alone. Precision is strong, but the systematic offset is large relative to the specification. Investigate and correct bias, then verify accuracy and repeatability again.
Practice case: Is a 35% Gage R&R always a gage failure?
Answer: It is generally unacceptable under common guidelines, but first check the study design. If all parts came from a very narrow range, PV may be understated. %Tolerance, graphs, sample representativeness, and the dominant variance component tell you whether the main problem is the system, the sample, or both.
Knowledge check
Check your understanding
Comprehension quiz
Pick the strongest answer, then submit. Explanations appear after you submit.
The Whole Lesson in One Minute
Signal
PV is the real difference among study parts. You want the gage to separate it clearly.
Random noise
Repeatability + reproducibility + interaction combine into Total Gage R&R.
Systematic error
Bias, linearity, and stability can make precise results consistently wrong.
- Define the measurement decision and choose the correct study.
- Use representative parts, normal appraisers, repeats, randomization, blinding, and repositioning.
- Inspect graphs before reducing the study to one number.
- Interpret %Study Variation, %Tolerance, %Contribution, and ndc as different views—not interchangeable labels.
- Match the dominant source to corrective action, then repeat the study under the improved method.
- Never claim accuracy, linearity, or stability from Gage R&R alone.
Lesson complete
You have moved from “What is Gage R&R?” to designing and diagnosing a full measurement study.
Authoritative Follow-Up
- Minitab: Crossed Gage R&R Study overview
- Minitab: Interpret Crossed Gage R&R results
- Minitab: Gage Linearity and Bias Study overview
- NIST/SEMATECH: Gauge R&R studies and measurement-process characterization
Use the applicable customer, industry, and organizational measurement-system requirements for acceptance decisions. Common 10%/30% bands are screening guidelines, not a replacement for risk-based criteria.