Concept study / Tests & research methodsConceptual illustration. Shapes, colors, and motion are not measured data.
Measurement uncertainty describes how well an estimate is known under its stated measurement model. It is not a confession that a result is useless. A repeated signal can be precise yet biased by calibration, and a highly variable signal can still support an average effect if its variation is characterized properly. NIST Technical Note 1297 provides a framework for combining uncertainty components. For the Hill program, every numerical claim should identify its measurand, unit, calibration, analysis choices and uncertainty. The aim is to determine which differences the experiment can resolve, not to attach decorative error bars after interpreting a plot.
Build an uncertainty budget
List contributions from instrument calibration, resolution, timing, environmental variation, alignment, data processing and model parameters. Record whether each contribution was estimated from repeated observations or other information. These classifications concern how uncertainty is evaluated; they do not map simply onto random versus systematic error. Document correlations. If two compared measurements share the same reference, some uncertainty may cancel in their difference while other components persist. A spreadsheet that adds every percentage in quadrature without examining dependencies can substantially misstate the uncertainty of the actual target quantity.
Propagate through the measurement model
State the function that converts raw readings into the reported effect. For small uncertainties and a sufficiently smooth function, sensitivity coefficients and covariance provide a local approximation. For nonlinear transformations, bounded parameters or strongly asymmetric distributions, simulation may be more appropriate. The proposed workflow stores each input distribution, its justification and the transformation code. Compare simulated output with a known benchmark before applying it to new data. If uncertainty in a background subtraction dominates the effect, improving sample count alone may do little to strengthen the result.
Report limits the reader can use
Publish the estimate, uncertainty convention, coverage assumptions and valid range. A confidence interval, credible interval and expanded measurement uncertainty answer different questions; identify which is being shown. Report a detection limit or upper bound when that is what the data support. For a claimed force, include the baseline and environmental channels. For a scaling exponent, include sensitivity to fitting range and selection. The uncertainty budget should help choose the next experiment: identify the largest reducible contribution rather than buying a more sensitive sensor without evidence that it addresses the limitation.
The mathematical lens
J is the sensitivity row vector of the measurement function and Σ the covariance matrix of inputs; this is a first-order approximation.
Turn it into a question.
- Question
- Is the predicted effect resolvable under the full uncertainty budget?
- Method
- Propagate calibration, noise and nuisance-parameter uncertainty before fixing sample size.
- What would count against it?
- If plausible systematic shifts exceed the effect, the current experiment cannot establish it.
An honest bound can be more informative than a dramatic central estimate.
Follow the original work.
These sources support the mechanisms and limits discussed here. They do not endorse Hill’s framework. Notes distinguish research proposals, reviews, and observations.
Continue the investigation
Edition 02 · 18 October 2026 · AI-assisted synthesis. See editorial method.