Glossary term

Measurement Bias

Engineering definition of measurement bias covering systematic error, mean error, reference comparison, bias correction, confidence interval and release evidence.

Branch: Engineering Physics
Glossary type: metric
Content: Glossary term
Updated: Jun 28, 2026
Revision: v1.0.1 · reviewed

Definition

metric

Measurement bias is the systematic difference between the expected measurement result and a reference value for the stated measurement condition.

Measurement bias is a systematic error component. It can come from calibration offset, gain error, installation effect, environmental coupling, algorithmic correction, reference mismatch, fixture loading or method definition. Bias is different from repeatability scatter: repeated readings can be tight and still biased.

Measurement bias is the systematic difference between a measurement result and a reference value for the stated condition. It is the part of error that tends to push results in one direction rather than scatter randomly around the reference.

Bias matters because a system can repeat very well and still be wrong. A sensor, inspection fixture, imaging algorithm, digital twin input or process analyzer may produce stable readings that are consistently high or low.

Error and Bias

For each comparison with a reference:

e_i=x_i-x_{ref,i}

where (x_i) is the measured value and (x_{ref,i}) is the reference value. A mean bias estimate is:

\displaystyle \bar{b}=\frac{1}{n}\sum_{i=1}^{n}e_i

A positive bias means the measurement tends to read high. A negative bias means it tends to read low.

Worked Bias Example

Five temperature comparisons against a reference produce errors:

0.6,\ 0.4,\ 0.7,\ 0.5,\ 0.8\ ^\circ\text{C}

The estimated mean bias is:

\bar{b}=0.60\ ^\circ\text{C}

The sample standard deviation of the errors is:

s_e=0.158\ ^\circ\text{C}

The data suggest a positive measurement bias, not just random scatter.

Confidence Interval

For a small sample, a confidence interval for mean bias can be screened as:

\displaystyle \bar{b}\pm t\frac{s_e}{\sqrt{n}}

Using (t=2.776) for a 95 percent interval with 4 degrees of freedom:

\displaystyle 0.60\pm2.776\frac{0.158}{\sqrt{5}}=0.60\pm0.20\ ^\circ\text{C}

The interval is approximately:

0.40\leq b\leq0.80\ ^\circ\text{C}

If the allowed mean bias is (0.5\ ^\circ\text{C}), this is not a clean pass.

Bias Correction

If the bias is stable, understood and valid over the range, a correction may be applied:

x_{corr}=x_{meas}-\hat{b}

Correction is only defensible when the reference basis, range, environment, configuration and uncertainty are documented. Correcting a bias found at one point may be unsafe if the error changes with range, temperature, loading, flow regime, image reconstruction setting or installation condition.

Bias Over Range

Bias should be checked at the points that matter for the decision. A zero offset test does not prove span accuracy. A mid-range reference does not prove behavior near saturation. For a simple linear screen, error can be separated into offset and gain terms:

e(x)=b_0+b_1x

where (b_0) is offset bias and (b_1) is gain-related bias. If (b_1) is significant, a single correction value may under-correct one end of the range and over-correct the other.

Bias Versus Repeatability

Repeatability describes scatter under unchanged conditions. Bias describes systematic offset relative to a reference. The two can coexist:

e_i=b+\epsilon_i

where (b) is systematic bias and (\epsilon_i) is random error. Averaging reduces random scatter in the mean, but it does not remove bias unless the bias is estimated and corrected.

Engineering Sources

Common sources of bias include zero offset, span error, reference mismatch, thermal gradient, pressure tap location, probe loading, fixture deflection, cable resistance, electromagnetic interference, algorithmic smoothing, clock offset, model-form error and human alignment.

In resistance-temperature measurements, uncompensated lead-wire mismatch can appear as a stable positive or negative temperature bias. That error may survive a bench calibration if the installed cable, terminal block, junction box or compensation wiring is different from the calibration setup.

In data assimilation and sensor fusion, shared bias is especially dangerous. Combining several biased measurements can make the uncertainty look small while the estimate remains wrong in the same direction.

Verification After Correction

After correction, engineers should verify residual error with independent points:

r_i=x_{corr,i}-x_{ref,i}

The correction is only useful if residuals are small enough for the tolerance, show no unmodeled trend and remain stable under the intended operating conditions. Otherwise the result belongs in the uncertainty budget or rejection decision, not as a hidden adjustment.

Evidence for Release

A useful bias review states the reference, traceability, comparison points, range, direction of approach, environmental condition, sample size, error data, mean bias, confidence interval, correction rule, residual uncertainty and whether the result affects past decisions.

For regulated, safety-critical or quality-critical use, the record should also say who owns the risk if bias correction is used near a limit.

Limits and Common Mistakes

Common mistakes include treating repeatability as accuracy, correcting bias without checking range dependence, using a weak reference as truth, hiding bias inside random uncertainty, averaging biased sensors, and ignoring shared calibration bias in multi-sensor estimates.

Another mistake is calling every offset a defect. Some bias can be corrected and controlled. The engineering question is whether the bias is known, stable, valid for the use case and small enough after correction and uncertainty to support the decision.

REF

Disciplines