A shift supervisor once walked into my office holding a shaft and a problem. The laboratory CMM measured the journal at 12.003 mm. The in-line tri-axial gauge measured it at 11.987 mm. The tolerance was ±20 microns. One system said the part was good; the other said it was scrap.

When your measurement systems disagree by 16 microns on a 40-micron tolerance, your entire quality infrastructure trembles. Statistical process control is useless if you cannot distinguish process variation from measurement noise. Your control plan becomes an opinion. Customer trust evaporates.

Gage correlation is the bridge between these two worlds. It is the systematic process of verifying that two or more measurement systems produce consistent, comparable results on the same parts.

MSA vs. Gage Correlation

Measurement Systems Analysis (MSA) tells you if a single gauge is capable—whether it is accurate, repeatable, and reproducible. Gage correlation goes further. It asks whether two capable systems agree with each other.

It is the difference between asking if a thermometer is accurate, and asking if two accurate thermometers show the same temperature.

In automotive and aerospace manufacturing, this is critical. You have a CMM in the lab measuring First Article. You have an in-line optical system measuring 100% of production. You have a supplier measuring before shipment, and a customer measuring at incoming inspection. Four measurement systems, one characteristic.

Where the calculation meets the shop floor: the gap between planned availability and the shift people actually work.
Where the calculation meets the shop floor: the gap between planned availability and the shift people actually work.

When Correlation Becomes Non-Negotiable

I have audited plants where engineers assumed that because two gauges were calibrated, they automatically agreed. Calibration proves traceability to a national standard; it does not prove alignment between different measurement physics.

Correlation is mandatory in specific scenarios. Production transfers between plants are a prime example. Without correlation, the exact same part can pass in Plant A and fail in Plant B. Introducing new measurement technology—replacing a contact probe with a laser scanner—breaks historical data continuity unless you correlate the old and new systems.

It is also the only way to resolve customer rejections. If a customer rejects parts based on their incoming inspection, and your final inspection says they are good, you cannot resolve the dispute without a pre-established correlation study. Standards like IATF 16949, VDA 6.3, and ISO/IEC 17025 expect demonstrable measurement consistency.

The Correlation Process

Correlation requires discipline, not advanced mathematics. The goal is to isolate the variation between the systems by tightly controlling every other variable. You must select critical characteristics—those flagged on the control plan as CC, SC, or CS.

Gage Correlation Methodology

  1. 01Define ScopeSelect critical characteristics (CC/SC) and map every gauge that measures them.
  2. 02Select PartsChoose minimum 10 parts spanning the full tolerance range, not just nominal.
  3. 03Randomise MeasurementMeasure parts in random order to eliminate thermal drift and operator fatigue.
  4. 04Analyse DataUse scatter plots and Bland-Altman to expose bias and proportional error.
  5. 05DocumentFile the report in the PPAP package and define periodic verification triggers.
Correlation is a sequence of controlled experiments designed to isolate measurement variance.

You need a minimum of 10 parts covering the entire tolerance range. Do not just pick nominal parts. Use three parts near the lower limit, four in the middle, and three near the upper limit. Every part must be dimensionally stable so it does not change physically during the test.

Measure each part on each system at least three times. Randomise the order. Measuring all parts sequentially on System A and then System B introduces systematic error due to temperature shifts or operator fatigue. The operators, the clamping method, and the procedure must remain identical. The only variable changing is the measurement system.

Data Analysis: Finding the Bias

Most engineers look at the average difference between two gauges and declare victory if it is small. This is dangerous. If the difference is -8 microns at the lower limit and +14 microns at the upper limit, the average might look acceptable, but the correlation is broken.

A scatter plot with a 45-degree reference line (y = x) provides a quick visual check. If points deviate systematically above or below the line, you have bias. If the points scatter widely, you have a precision problem.

Bland-Altman analysis is better for industrial use. Plot the mean of both measurements on the X-axis and the difference between them on the Y-axis. Add limits of agreement (mean ± 1.96 × standard deviation). This immediately reveals proportional bias—where differences grow as the measurement size increases.

Linear regression provides the quantitative metric. Look for an R² greater than 0.95, a slope close to 1.0, and a y-intercept near zero. If the slope is 0.98 and the intercept is 0.005 mm, you have near-perfect correlation.

Calibration proves traceability to a standard; it does not prove alignment between different measurement physics.

Setting Acceptance Criteria

Correlation is not a binary pass-or-fail. You need quantitative criteria. My rule is that the maximum difference between systems must not exceed 20% of the tolerance band. On a 40-micron total tolerance, the maximum allowable deviation is 8 microns.

Correlation Acceptance Thresholds

20%Standard limitMax gauge-to-gauge deviation as percentage of total tolerance band.
10%Safety criticalMandatory ceiling for CC/SC characteristics affecting regulatory compliance.
0.95R² floorMinimum coefficient of determination for linear regression validation.
Tighter tolerances are required for safety and regulatory critical characteristics.

For safety or regulatory critical characteristics, I reduce this to 10% of the tolerance band. If the correlation fails these criteria, the real engineering work begins.

Diagnosing the Root Cause

In the case of the disagreeing CMM and tri-axial gauge, the scatter plot revealed an S-curve rather than a straight line. At nominal dimensions, the correlation was perfect. At the extreme limits, the differences spiked. Bland-Altman confirmed a proportional bias.

The root cause was fixturing. The CMM used a pneumatic clamp that caused microscopic deformation. On nominal parts, the deformation was negligible. On parts at the edge of the tolerance, where the geometry was already stressed, the clamping force distorted the part by an additional 10 to 12 microns. The in-line tri-axial system used gravitational seating—no lateral force, no deformation.

The solution was lowering the CMM pneumatic pressure and adding support points aligned with the GD&T datum scheme. Post-correction, the systems aligned within 4 microns—well inside the acceptance window. Document these fixes and update the PPAP package.

Industry 4.0 and Continuous Correlation

Gage correlation is not a one-time milestone. Gauges degrade, calibrations drift, and operators change. Periodic verification—minimally annually or during re-qualification—is mandatory.

Modern IoT connectivity allows for automated correlation. I have implemented dashboards that hourly compare in-line gauge data against a reference CMM. When correlation drops below a threshold, the system stops part approval and alerts engineering. Within the first month, this automated check caught a slow degradation in an optical sensor—a drift that would have gone unnoticed until the next quarterly calibration.

The biggest quality failures rarely start with a bad process. They start with flawed measurement of that process. Gage correlation ensures that what you measure is what you actually have.