Gage Bias and Linearity Studies

Gage bias is the difference between the average of a gage's readings and a traceable reference value for the same feature; linearity is how much that bias changes across the gage's measuring range. A bias-and-linearity study measures both, so you know not just whether a gage reads high or low, but whether it reads high in one place and low in another.

Repeatability tells you whether a gage agrees with itself. Bias and linearity tell you whether it agrees with the truth. A caliper can be perfectly consistent and still be wrong by two thousandths on every part, quietly shifting your whole process off center. And a gage that reads dead-on at the low end of its range can drift off at the high end, so a part that passes at 10 millimeters and a part that passes at 100 millimeters are being judged by two different rulers. This guide covers what bias and linearity are, how to run a study using reference standards, how to read the numbers, and where the two errors fit alongside gage R&R in a full measurement systems analysis.

What is gage bias?

Gage bias is a systematic, one-direction error: the gage's average reading sits a fixed amount above or below the true value. If you measure a reference block whose certified size is 25.000 millimeters and your gage averages 25.008 across many readings, the bias is +0.008. It is not random scatter, it is a consistent offset that pushes every measurement the same way, which is exactly why it is dangerous. Random noise averages out; bias does not.

The reference definition, from the AIAG Measurement Systems Analysis manual, is that bias is the difference between the observed average of measurements and a reference value, where the reference value is established by a more accurate method traceable to a national standard. Because bias moves every reading in the same direction, it eats into your effective tolerance from one side. A part that is truly at the edge of the low limit will read as failing if the gage is biased low, and a part that is truly over the high limit can read as passing if the gage is biased high. You scrap good parts, ship bad ones, or both.

What is gage linearity?

Gage linearity is how the bias changes across the gage's operating range. A gage has good linearity when its bias is about the same size everywhere, small and steady from the bottom of the range to the top. It has poor linearity when the bias grows, shrinks, or flips sign as the measured value gets larger. The classic example: a scale that reads one pound light on a 150-pound load but five pounds light on a 200-pound load. Same instrument, very different error depending on where you are on the dial.

You study linearity by measuring several reference parts spread across the range, computing the bias at each one, and plotting bias against the reference value. If the points scatter randomly around zero, linearity is good. If they fall along a sloped line, the bias is changing with size and the slope tells you how fast. A steep slope means you cannot apply a single correction factor to the whole range, the error you would subtract at the low end is not the error you have at the high end.

How do you run a gage bias and linearity study?

You need reference parts whose true sizes are known and traceable, spanning the range you actually measure. The AIAG approach is to pick several parts across the operating range, get a trusted reference value for each, then have an operator measure each part many times under normal conditions. Here is the procedure:

1. Choose reference parts across the range. Pick at least five parts spread from the low end to the high end of the gage's operating range. Two parts only give you a straight line by definition and hide any curve; more parts across the span reveal the real shape of the error.
2. Establish a traceable reference value for each. Measure each reference part with a higher-accuracy method, a calibration lab, a CMM, or certified gage blocks, traceable to a national standard. This is the truth you will compare the working gage against.
3. Measure each part repeatedly with the working gage. Have one operator measure every reference part about 10 or more times under normal shop conditions, in random order, using the gage as it is really used on the floor.
4. Compute the bias at each part. For each reference part, average the working-gage readings and subtract the reference value. That difference is the bias at that point on the range.
5. Test whether the bias is significant. Check whether each bias is statistically different from zero. A small bias inside the noise may not be worth correcting; a bias that clearly clears the noise needs action.
6. Plot bias against reference value and fit a line. Draw the bias-versus-reference plot, fit a best-fit line, and read the slope. The slope is your linearity; the scatter around the line and the intercept round out the picture.
7. Judge and act. Decide whether bias and linearity are acceptable for the tolerance you are gauging. If not, calibrate, adjust, or replace before you trust the gage for accept/reject decisions.

By the numbers. The definitions and study design here follow the AIAG Measurement Systems Analysis reference manual, which treats bias, linearity, and stability as the "location" errors of a measurement system and repeatability and reproducibility as the "width" errors (AIAG, Measurement Systems Analysis). AIAG's method calls for choosing reference parts across the operating range and taking on the order of ten readings per part to separate bias from ordinary repeatability. The underlying variance and regression math, averaging readings, testing whether a bias differs from zero, and fitting the linearity line, is laid out in the NIST/SEMATECH engineering statistics handbook's measurement process characterization section (NIST/SEMATECH e-Handbook). Both are recognized primary references; neither invents a single acceptance number that fits every part, because the acceptable bias always depends on the tolerance you are gauging.

How do you read the bias and linearity numbers?

Read bias as a fraction of the tolerance, not as a raw number. A bias of 0.002 is trivial on a part with a 0.050 tolerance and disqualifying on a part with a 0.004 tolerance. The common way to express it is percent bias, the bias divided by the tolerance (or by the process variation), so it scales with what you are measuring. A bias that eats a meaningful slice of the tolerance means the gage is stealing usable tolerance from you and pushing accept/reject decisions off center.

Read linearity as the slope of the bias line and the range of bias across the span. A near-flat line means one correction, or one calibration offset, works everywhere. A steep line means the gage behaves differently at different sizes, and you either need it serviced or you need to restrict it to the part of the range where its bias is acceptable. If the study also shows the bias at each point is statistically inseparable from zero, you may have a gage that is fine as-is; if the bias clears the noise and grows with size, you have both a bias problem and a linearity problem to fix.

What causes bias and non-linear error?

Bias usually comes from calibration and setup. The gage was zeroed against the wrong master, the master itself has drifted, the temperature is off from the standard 20 degrees Celsius so the part has expanded, or the operator applies more measuring force than the calibration assumed. These shift every reading the same way. A gage that has never been calibrated against a traceable standard is almost guaranteed to carry some bias you simply cannot see.

Non-linear error tends to come from the mechanism itself: worn threads on a micrometer that bind more at one end of travel, a dial indicator whose gear train is tighter in part of its sweep, a lever probe whose geometry adds cosine error as the angle grows, or a sensor whose response curve bends at the extremes of its range. This is why linearity has to be checked across the range and not at a single convenient point, a gage calibrated only at mid-scale can hide a large error at the ends. Confirming the gage's discrimination is fine enough for the tolerance, roughly one-tenth of it, keeps you from chasing "bias" that is really just a gage too coarse to see the feature.

How do bias and linearity fit with gage R&R and MSA?

Bias and linearity are the accuracy half of a measurement system; repeatability and reproducibility are the precision half. A full measurement systems analysis covers both, because a gage can pass one and fail the other. A gage R&R study can come back clean, tight, consistent, operators agreeing, while the gage is biased two thousandths low on every part, because gage R&R measures variation, not accuracy against a truth. That is the trap: repeatability makes a biased gage look confident. You need reference standards, which is what a bias-and-linearity study adds, to catch an offset that R&R is structurally blind to.

In practice the order is: confirm the gage's resolution is fine enough, run bias and linearity against traceable references to fix accuracy, run gage R&R to nail down precision, then hold the gains with a gage stability study that watches for drift over time. Only then do the numbers feed cleanly into statistical process control and process capability both of which assume the measurement is telling the truth. A biased gage inflates or deflates a Cpk and shifts control-chart center lines, so an untrustworthy gage can make a good process look bad or a drifting process look fine.

The catch with bias and linearity is that they are point-in-time studies, and a gage that is accurate today can drift tomorrow. That is where live data on the floor pays off. When measurement results and out-of-tolerance events are captured at the point of inspection instead of on a clipboard, a gage that is starting to read consistently high shows up as a pattern long before the next calibration cycle catches it. That live feedback is what Harmony gives a plant, and it fits the broader discipline of good quality work: decisions are only as good as the data, and the data is only as good as the gage. CLS made that shift, from measurements found the next morning to measurements visible during the shift, which is what keeps a calibrated gage from quietly going out of true between studies. You can see how that works alongside the rest of the plant on the features overview.