Measurement Error, Systematic Bias & Uncertainty Budgets

A numerical reading is not yet a measurement result. One must define the measurand—the quantity intended to be measured—specify the conditions and model connecting observations to it, apply justified corrections, and state the uncertainty that remains. “The voltage is 10 V” is incomplete if bandwidth, location, waveform convention, loading, calibration, and relevant environment are unspecified.

Error and uncertainty are different concepts #

Measurement error is the difference between a measured value and a reference quantity value. The exact error is ordinarily unknown because the true quantity value is unknown. Uncertainty describes the dispersion of values that could reasonably be attributed to the measurand, given the model and available information.

This distinction prevents two common mistakes:

• A correction for known bias changes the estimate; its imperfect value contributes uncertainty.
• Repeated readings can estimate scatter but do not reveal a constant calibration bias shared by every reading.

“Random” and “systematic” describe how effects behave across repetitions or conditions. Type A and Type B describe how uncertainty is evaluated. Type A uses statistical analysis of observations; Type B uses other information such as calibration certificates, specifications, resolution, prior data, or physical bounds. A systematic effect can have a Type A evaluation, and a randomly varying effect can be evaluated from prior Type B information.

Build a measurement model #

Represent the output quantity as

$$Y=f(X_1,\ldots,X_m),$$

where the $X_i$ include direct readings, calibration corrections, environmental influences, geometry, constants, and any other inputs that materially affect the measurand. Each input estimate $x_i$ has a standard uncertainty $u(x_i)$ and may be correlated with other inputs.

For a sufficiently linear model near the estimates, the combined standard uncertainty is

$$u_c^2(y)=\sum_i c_i^2u^2(x_i)+2\sum_{i<j}c_ic_j u(x_i,x_j), \qquad c_i=\left.\frac{\partial f}{\partial X_i}\right|_{x}.$$

The sensitivity coefficient $c_i$ converts uncertainty in input units into output units. The covariance terms are load-bearing: a shared reference can make errors move together, so treating inputs as independent may either understate or overstate uncertainty.

flowchart TD
  D["Define measurand and conditions"] --> M["Write measurement model"]
  M --> I["Estimate inputs and correlations"]
  I --> P["Propagate distributions or covariance"]
  P --> R["Report estimate, uncertainty, coverage and traceability"]

For strong nonlinearity, bounded inputs near a limit, asymmetric distributions, or discontinuous decision rules, first-order propagation can be misleading. Propagating input distributions numerically is then more faithful, but it does not repair an incomplete model.

Worked uncertainty budget #

Power is computed from independent voltage and current measurements:

$$P=VI,$$

with $V=10.00\ \mathrm{V}$, $u(V)=0.05\ \mathrm{V}$, $I=2.00\ \mathrm{A}$, and $u(I)=0.02\ \mathrm{A}$. The estimate is $P=20.00\ \mathrm{W}$. Sensitivities are $\partial P/\partial V=I$ and $\partial P/\partial I=V$, so

$$u_c(P)=\sqrt{(2.00\cdot0.05)^2+(10.00\cdot0.02)^2}=0.224\ \mathrm{W}.$$

An expanded uncertainty $U=ku_c$ with $k=2$ is about $0.45\ \mathrm{W}$, so one could report $P=(20.00\pm0.45)\ \mathrm{W}$ with the stated coverage convention and assumptions. The current term contributes four times as much variance as the voltage term: improving voltage accuracy alone has little leverage.

If voltage and current came from a shared gain reference, add their covariance. If the quoted ± limits are rectangular bounds rather than standard deviations, convert them according to the justified input distribution before propagation. Never silently treat every specification limit as one standard deviation.

What belongs in an uncertainty budget #

Traceability is not a brand name or an unbroken stack of stickers. It is a documented chain of calibrations, each contributing uncertainty, connecting a result to a stated reference.

Bias, correction, and validation #

If an instrument reads $+0.30$ units on a reference with an estimated correction uncertainty of $0.05$, apply a $-0.30$ correction when the model and conditions justify it, then include $0.05$ and other reference uncertainties in the budget. Leaving the known effect uncorrected and merely widening an interval loses information and can leave the estimate systematically wrong.

Validation asks whether the model contains all relevant effects. Vary conditions, compare independent methods, measure reference materials, reverse polarity or geometry where meaningful, and inspect residuals against time and environment. Agreement between repeated readings tests repeatability, not accuracy.

Reporting without false precision #

A complete result states the estimate, unit, measurand definition, combined or expanded uncertainty, coverage factor or probability interpretation, important assumptions, and traceability. Round uncertainty to a sensible number of significant digits, then round the estimate to the same decimal place.

Uncertainty is not a mistake bar and not a guarantee that the true value lies inside a symmetric interval. It is the quantified consequence of a measurement model. The most dangerous budget is a precise calculation around the wrong measurand or a missing systematic effect.