Key ideas
- The result of a measurement may depend on several input quantities, corrections, and a model that represents the process.
- NIST represents contributions using standard uncertainties and distinguishes Type A and Type B evaluations by how they are estimated.
- Combined uncertainty alone does not guarantee a normal distribution or a confidence interval: those interpretations depend on specific statistical conditions.
From the measurement question to the result
A measurement does not always obtain the quantity of interest directly. NIST calls this quantity the measurand and explains that, in some cases, it is determined from other quantities through a functional relationship, or measurement equation. The estimated result then depends on the input estimates and on whether the model represents the measurement process. This approach is not merely a physical relationship: it must include the quantities that may contribute significantly to the uncertainty of the result.
The example provided by NIST is the power dissipated by a resistor whose resistance changes with temperature. The power depends on the potential difference, the resistance at a reference temperature, the temperature coefficient, and the temperature. The equation presented is \(P = f(V, R_0, b, t) = V^2/R_0[1 + b(t - t_0)]\). Thus, uncertainty in the estimate of the power arises from the uncertainties associated with those input quantities. The example shows why it is useful to ask not only “what value was measured?” but also “what inputs and assumptions does that value depend on?”.
The explanation concerns measurement uncertainty, not a universal assessment of the quality of all scientific evidence. Its scope is metrological: it provides a framework for representing and expressing uncertainties associated with measurement results. The NIST page presents its “essential elements” as a summary of a widely adopted method adapted from NIST Technical Note 1297, which is itself based on the International Organization for Standardization guide on expressing uncertainty in measurement.
Identifying and representing contributions
According to NIST, input quantities may include corrections and correction factors, as well as sources of variability such as different observers, instruments, samples, laboratories, or times of observation. The model should incorporate the relevant significant contributions; omitting a relevant source may leave the representation of the process incomplete. In practice, this makes it useful to state explicitly what was measured directly, what was estimated, and which factors were included in the equation.
NIST represents each uncertainty component by a standard uncertainty, expressed as an estimated standard deviation. For a Type A evaluation, the component is represented by a statistically estimated standard deviation associated with a statistically estimated variance. For a Type B evaluation, the standard uncertainty is estimated from the available information and an assumed probability distribution. Both categories describe evaluation methods, not a hierarchy in which one is automatically better than the other.
| Aspect | Type A | Type B |
|---|---|---|
| Representation described by NIST | Statistically estimated standard deviation | Estimate based on available information and an assumed distribution |
| What it does not imply by itself | That all relevant sources have been considered | That the estimate is less valid than a Type A estimate |
The table summarizes the methodological distinction described by NIST; it does not determine which evaluation applies to a particular case. That decision depends on the available information and the measurement procedure.
Combining does not eliminate uncertainty
Combined standard uncertainty, whose recommended symbol is \(u_c\), represents the estimated standard deviation of the result. NIST states that it is obtained by combining the individual standard uncertainties and, where applicable, the covariances, through propagation of uncertainty. The combination considers components evaluated by both Type A and Type B methods. Therefore, bringing the contributions together into a combined value does not erase their origins: it depends on the input estimates and on how they relate to one another.
The source also distinguishes error from uncertainty. The uncertainty associated with a correction applied to compensate for a systematic effect is not the systematic error caused by that effect; it represents uncertainty due to incomplete knowledge of the required correction value. NIST recommends applying corrections for recognized significant systematic effects, although it notes that in some cases it may not be practical to do so because of resource limitations. The expression of uncertainty should then conform to its guidelines as far as possible.
This distinction matters when communicating results: a correction does not by itself make a measurement exact, and the uncertainty associated with it should not be described as if it were the error itself. A careful conclusion identifies the recognized effect, states whether a correction was applied, and retains the relevant uncertainty, rather than turning it into a claim of absolute certainty.
What a combined uncertainty allows us to conclude
In practical situations, the probability distribution characterized by the result and its combined standard uncertainty may be approximately normal. NIST relates this possibility to specific conditions: for example, that the estimate depends on several input quantities with suitable distributions, that their uncertainties contribute in comparable magnitudes, and that the linear approximation used in propagation is adequate. A normal distribution should not be assumed just because \(u_c\) has been calculated.
When the stated conditions are met and the combined uncertainty itself has negligible uncertainty, NIST describes the interval from \(y-u_c(y)\) to \(y+u_c(y)\) as associated with an approximate confidence level of 68 %. The source also cautions that “confidence interval” has a specific statistical meaning and that not every interval constructed from \(u_c\) has that meaning. In its guidelines, the interval is interpreted through the fraction of the probability distribution it covers, called the coverage probability or confidence level of the interval.
Consequently, an uncertainty figure cannot be interpreted independently of its method, model, and assumptions. To read or communicate a result, it is useful to ask what measurand was estimated, which inputs and corrections were included, how the contributions were evaluated, and what conditions justify interpreting the interval. These questions delimit what the result allows one to claim; they do not replace a specific evaluation of the procedure or, by themselves, prove the validity of a broader scientific conclusion.
References
What we know and its limits
The method described concerns the evaluation and expression of uncertainty in measurements. Interpreting an interval based on combined standard uncertainty depends on statistical assumptions and conditions; it does not determine whether these are met in a particular measurement. It is not a general guide to evaluating every kind of scientific evidence. Box plots summarize and compare the location and variation of data, but they do not constitute an evaluation of measurement uncertainty or corroborate the metrological claims described.
Original sources3
The original sources for exploring this explanation and its limits.
- Box Plots: Center, Spread, and Outliers NIST
-
Measurement uncertainty
NIST
Institutional reference for context. It does not constitute independent confirmation of the main source's findings. - How the different contributions to uncertainty are combined NIST
Image credits
Flat lay of laboratory glassware including beakers and flasks on a neutral surface. · Photo by Tara Winstead on Pexels · Original source · Pexels License · Resized and re-encoded as JPEG; full framing preserved.