Process Capability and Gage R&R: Can You Trust Your Cpk?

A process capability calculator compares measured output with specification limits. But the measurements contain two signals: differences between parts and variation introduced by the measurement system. Before treating a low Cpk as a machine problem, investigate how much of the observed spread comes from measurement.

Gage repeatability and reproducibility, usually called Gage R&R, helps investigate that question. This guide connects measurement system analysis with capability analysis through a numerical example and a practical study sequence.

What capability analysis and Gage R&R tell you

Capability asks whether a stable process can meet specifications. For a normal-distribution model with two specification limits, the familiar population formulas are:

Cp = (USL - LSL) / (6 * sigma)
Cpk = min((USL - mean) / (3 * sigma), (mean - LSL) / (3 * sigma))

Cp describes the tolerance width relative to process spread. Cpk also accounts for the position of the mean. Specifications come from product requirements; control limits describe process behavior. They serve different purposes. See the NIST introduction to process capability.

Gage R&R examines measurement consistency. Repeatability concerns repeated measurements of the same characteristic by the same appraiser using the same instrument. Reproducibility concerns differences associated with appraisers. A crossed study has each appraiser measure the same parts repeatedly. The ASQ guide to Gage R&R explains these concepts and alternative study structures.

Capability evaluates output against requirements. Gage R&R evaluates whether the measurement process can distinguish differences relevant to that evaluation.

Why measurement variation changes a capability result

Under an additive, independent, zero-mean measurement-error model:

Observed measurement = true part value + measurement error
Observed variance = part variance + measurement variance
sigma_observed = sqrt(sigma_part^2 + sigma_measurement^2)

The variances add; the standard deviations do not. A formal crossed model can also include an operator-by-part interaction. See the SAS description of the crossed Gage R&R variance model.

Extra random measurement variation widens the observed distribution. With the same mean and specification limits, that reduces the calculated capability indices. A systematic bias is a separate problem: it moves the measured mean and can make the reported result look better or worse.

Worked example: a dimension that looks less capable than it is

Consider a synthetic dimensional example with these assumed population quantities:

• Lower specification limit: 9.70 mm.
• Upper specification limit: 10.30 mm.
• Process mean: 10.04 mm.
• True part-to-part standard deviation: 0.08 mm.
• Measurement-system standard deviation: 0.04 mm.

These are teaching assumptions, not estimates produced by a raw-data Gage R&R study. Assume the process and measurement system are stable, the error is unbiased and independent, and the normal capability model is appropriate.

The observed standard deviation is:

sqrt(0.08^2 + 0.04^2) = 0.089443 mm

Using the observed spread gives:

Cp = 0.60 / (6 * 0.089443) = 1.1180
Cpk = min(0.26, 0.34) / (3 * 0.089443) = 0.9690

Using the assumed true part spread instead gives Cp = 1.2500 and Cpk = 1.0833. Measurement noise explains the gap between those calculations. It does not establish that the underlying process satisfies the customer's acceptance criteria.

If the team improves the gauge while leaving production untouched, the measured capability may rise. That is a better measurement of the process, not evidence that the parts themselves improved.

Why subtracting a Gage R&R percentage from Cpk is wrong

Cpk is a ratio involving standard deviation and distance to a specification limit. A Gage R&R percentage has a different denominator. There is no valid shortcut such as corrected Cpk = Cpk + %GRR.

Even variance subtraction requires a justified model and compatible variance estimates. It should not become a way to replace required measured capability results with a more favorable number. Preserve the original result and document any separate measurement-error analysis.

Three Gage R&R percentages that answer different questions

For the same example, using six standard deviations as the study width:

• Percent study variation: 100 * 0.04 / 0.089443 = 44.72%. This compares measurement spread with total observed spread.
• Percent contribution: 100 * 0.04^2 / 0.089443^2 = 20.00%. This compares measurement variance with total observed variance.
• Percent tolerance: 100 * 6 * 0.04 / 0.60 = 40.00%. This compares measurement study width with the specification width.

The 44.72% and 20.00% figures describe the same variance components on different scales. They are not contradictory. Keep the denominator visible when reporting a percentage; see the Minitab guide to crossed Gage R&R results.

A study using deliberately broad part variation can produce a different percent study variation than one restricted to a narrow production stream. Document part selection and the intended decision. Apply the acceptance criteria required by your customer or quality system to the appropriate metric.

Plan two datasets, not one combined shortcut

The measurement study and capability study have different sampling purposes.

For a nondestructive crossed measurement study, select parts that represent the measurement challenge, have the relevant appraisers measure every part repeatedly, randomize the measurement sequence and conceal previous readings. Record part, appraiser, trial and measured value in separate columns. If testing destroys or changes the part, choose a study design that accounts for that constraint.

For capability, collect a separate, representative production sample across the conditions and time period you intend to describe. Preserve time order and any rational subgroup structure. Repeated gauge readings of the same few selected parts do not become independent production observations simply because the spreadsheet contains many rows.

Check bias, resolution and stability as well as repeatability and reproducibility. NIST's measurement-process framework treats these as distinct issues.

Use the workflow in Weibull.app

Open Weibull.app and use Gage R&R with the part, appraiser, trial and measurement columns. Use its current summary as an exploratory view of equipment, appraiser and part variation. It does not fit the full crossed ANOVA variance-component model or estimate operator-by-part interaction; use a validated method matching the required study protocol for formal acceptance.

Then use Process Capability on the representative production measurements with the agreed LSL and USL. Review the mean, standard deviation and indices together.

The current capability calculation uses the overall sample standard deviation of the selected measurements. Its displayed Cp/Cpk labels do not establish that a within-subgroup estimator was used. Record that basis when comparing results with reports that distinguish within-based Cp/Cpk from overall-based Pp/Ppk.

Frequently asked questions

Does a good Gage R&R result prove the process is capable?

No. A consistent measurement system may accurately reveal a process with excessive spread or an off-center mean. Compare the production output with its specifications separately.

Can a good Cpk hide a measurement problem?

Yes. For example, bias can move the measured mean closer to the specification midpoint, while coarse resolution can conceal real differences. Review the measurement evidence rather than relying on the index alone.

What should an improvement report retain?

Keep the measurement-study method, part-selection rationale, capability sampling period, specification revision, standard-deviation estimator and the action being evaluated. That allows reviewers to distinguish better production from better measurement.