Does a low repeatability standard deviation prove a measurement is correct?

No. A stable bias can coexist with highly repeatable readings.

Can the assessment include software?

Yes. If software transforms sensor inputs into the reported measurement, it is part of the measurement system being assessed.

AI Glossary / Statistics and probability / Intermediate

MCA

Definition

MCA means measurement capability assessment, also called measurement capability analysis, in the supplied NIST statistical glossary. It examines whether a measurement system can provide results suitable for its intended purpose. Core topics include repeatability and reproducibility; a broader assessment also considers bias, stability, resolution, sensitivity, and unwanted responses to changing measurement conditions.

What it means

Before using measurements to control a process or train a model, a team needs to know what variation comes from the thing being measured and what comes from the measurement system itself.
An instrument can give closely repeated readings while consistently reporting the wrong level. It can also be well calibrated on one reference item yet perform poorly across operators, operating ranges, or environmental conditions. Capability therefore concerns the measurement process in its actual use, not just the instrument's advertised specification.

How it works

Define the measurand: the quantity intended to be measured. Specify the operating range, units, decision requirements, and conditions of use. A system adequate for detecting a large process shift may be inadequate for deciding whether a value lies just inside a narrow tolerance.
Map the measurement procedure, including instrument, software, operator, preparation, environment, and reference standards. Differences in positioning, lighting, temperature, or timing can be part of the system's performance rather than irrelevant background noise.
Design repeated measurements that separate relevant sources of variability. Repeatability concerns variation under specified repeated conditions. Reproducibility examines variation when specified conditions change, such as operators or instruments. State exactly which conditions change in the study.
Use representative items across the range of interest. Measuring only a convenient reference object can conceal problems near boundaries or for difficult geometries. Randomize or otherwise control measurement order where appropriate so drift is not confused with item or operator effects.
Analyze variability with a model appropriate to the study design. In a simple variance-components model, independent contributions can be combined by adding variances and taking the square root. Interactions or correlated sources require more careful treatment.
Assess bias using suitable reference values and their uncertainty. A consistent offset is different from random repeatability variation. Investigate whether bias changes across the measurement range, which can indicate a linearity or calibration problem.
Check stability over time and sensitivity to intended and unintended changes. A system should detect relevant changes in the measurand without responding excessively to irrelevant features. A camera-based measurement system, for example, may respond to lighting even when the physical dimension is unchanged.
Compare results with an explicitly stated requirement. Ratios involving process variation, tolerance width, or uncertainty answer different questions and must not be mixed. A percentage alone is not a complete capability conclusion, and acceptance criteria should follow the application rather than an unexplained universal cutoff.
Document limitations and monitor the system after changes. Replacing a sensor, updating analysis software, changing operators, or using a new range of items can invalidate assumptions from the original assessment. An MCA is evidence about a defined configuration and purpose, not permanent certification of every future use.

Three practical examples

A production team checks whether a gauge can distinguish meaningful differences between manufactured parts.
A laboratory studies variation across operators and days before comparing small changes in a measured quantity.
An AI vision team assesses a camera-and-model measurement pipeline across lighting conditions, item positions, and relevant sizes.

Think of it this way

Think of checking the ruler before arguing about the lengths of objects. You need to know whether repeated readings agree, whether different people read it similarly, and whether the markings match a trustworthy reference. A consistent but stretched ruler can still mislead every downstream calculation.

Worked example

An original illustration assumes a fitted measurement study has estimated two independent random standard-deviation components: 0.03 units for repeatability and 0.04 units for the additional reproducibility component. No interaction is included in this simplified model.
The combined random measurement standard deviation is sqrt(0.03^2 + 0.04^2) = sqrt(0.0025) = 0.05 units. Adding the standard deviations directly would give 0.07 and would not match the assumed independent variance-components model.
Suppose the application defines a six-standard-deviation study spread and has a specification interval of width 2 units. The calculated spread is 6*0.05 = 0.30 units, or 100*0.30/2 = 15% of that width. This is a stated comparison convention, not a universal probability or acceptance rule.
Separately, a reference item with assigned value 50.00 has a mean measured value of 50.08. The observed difference is +0.08 units. That possible bias is not explained away by the 15% spread ratio. Its interpretation also needs reference uncertainty, sampling uncertainty, and checks across the operating range.

Illustrative case study

In a fictional case, Alder Components plans to replace a manual dimensional check with an AI vision system. Repeated images under fixed conditions produce similar estimates, but the team expands the assessment before adoption.
Measurements across lighting settings reveal a systematic shift for reflective parts. Engineers revise the imaging procedure, repeat the relevant assessment, and document the tested range. The case illustrates why repeatability alone is insufficient; it does not report a real system's accuracy or imply that one successful trial validates every material and configuration.

Common misconceptions and mistakes

Do not equate repeatability with accuracy or treat calibration as a complete capability assessment. Do not compare a variance percentage directly with a standard-deviation percentage.
MCA is an ambiguous acronym. In this entry it means measurement capability assessment, following the source glossary; it does not mean multiple correspondence analysis. Also distinguish measurement capability from the capability of the production process being measured.

Business use

An MCA helps prevent unreliable measurements from driving product rejection, process adjustments, or model training. It can show whether the main improvement should target equipment, procedures, operating conditions, or data processing. Define the required decision precision before purchasing more sophisticated technology; a finer display resolution does not automatically produce more trustworthy measurements.

Check your understanding

Why are the example's 0.03 and 0.04 components combined as 0.05?

Answer: The stated model assumes independent random contributions, so their variances add.

Does the resulting 15% prove suitability?

Answer: No; the requirement, bias, stability, and other relevant uncertainties still need assessment.

Related terms

Repeatability
Bias
Calibration
Metrology study
Measurement system analysis
Gauge R and R
Reproducibility
Measurement uncertainty

Learn next

Next, study crossed and nested gauge studies and variance components. Then connect their results to an uncertainty budget and to the specific acceptance decisions the measurements will support.

Variance components

Sources

NIST/SEMATECH statistical glossary, MCA: https://www.itl.nist.gov/div898/handbook/glossary.htm NIST/SEMATECH e-Handbook, Gauge R and R studies: https://www.itl.nist.gov/div898/handbook/mpc/section4/mpc4.htm NIST/SEMATECH e-Handbook, Important issues for a gauge study: https://www.itl.nist.gov/div898/handbook/mpc/section4/mpc41.htm NIST/SEMATECH e-Handbook, Uncertainty analysis: https://www.itl.nist.gov/div898/handbook/mpc/section5/mpc5.htm

Previous: MBR Next: MCS algorithm

MCA