Uncertainty and Error in Scientific Measurements

Every measurement ever made contains an error — not in the sense of a mistake, but in the mathematical sense that no instrument, however precise, captures reality perfectly. Uncertainty and error are not failures of science; they are part of its architecture. This page covers what those terms mean formally, how systematic and random errors behave differently, where they appear in real laboratory and field settings, and how scientists decide when a measurement is good enough to act on.

Definition and scope

A measurement result is only half the information. The other half is the uncertainty attached to it — the quantified range within which the true value is expected to fall. The National Institute of Standards and Technology (NIST) defines measurement uncertainty formally in its NIST Technical Note 1297 as "a parameter, associated with the result of a measurement, that characterizes the dispersion of the values that could reasonably be attributed to the measurand." That definition, drawn from the internationally adopted Guide to the Expression of Uncertainty in Measurement (GUM), published by the Joint Committee for Guides in Metrology (JCGM 100:2008), is the working standard across metrology labs worldwide.

Error and uncertainty are related but not identical. Error refers to the difference between a measured value and the true value. Uncertainty is the statistical envelope around that difference — an acknowledgment that the true value is, in practice, unknowable with infinite precision. The scope of these concepts runs from undergraduate physics labs measuring the acceleration due to gravity (approximately 9.81 m/s²) to clinical trials measuring drug efficacy and satellite telemetry tracking orbital decay.

How it works

The GUM framework distinguishes two categories of uncertainty evaluation, and the distinction matters enormously for how data get interpreted.

Type A uncertainty is evaluated by statistical methods — calculating a standard deviation from repeated measurements. If a scale weighs a reference mass 30 times and produces a distribution of readings, the standard deviation of that sample characterizes the Type A uncertainty. It shrinks as more measurements are taken, following a 1/√n relationship.

Type B uncertainty is evaluated by means other than statistics — calibration certificates, manufacturer specifications, physical reasoning, or published reference data. A thermometer calibrated against a NIST traceable reference carries a stated uncertainty from its calibration certificate; that figure feeds directly into the Type B budget of any experiment using that thermometer. Type B does not improve with repetition, because it reflects fixed limitations of the instrument or reference.

These two types are combined into a combined standard uncertainty by summing their squares and taking the square root — a process formally called quadrature addition. The result is then multiplied by a coverage factor (typically k = 2 for approximately 95% confidence) to produce the expanded uncertainty reported alongside a result.

Separately from uncertainty categories, errors divide into systematic and random. Systematic errors shift all readings in the same direction — a miscalibrated zero point, for example, adds a constant offset to every measurement. Random errors scatter readings unpredictably above and below the true value. Averaging reduces random error but leaves systematic error untouched, which is why calibration cannot be treated as optional. This distinction is foundational to the broader picture of how science works as a conceptual process.

Common scenarios

Error and uncertainty show up in recognizably different forms depending on context.

1. Instrument resolution limits. A digital multimeter displaying three significant figures introduces a quantization uncertainty of ±0.5 in the last digit by definition. This is a Type B contribution.

2. Environmental interference. Temperature fluctuations alter the resistance of precision resistors, introducing a time-varying systematic drift. Metrology labs maintain temperature to within ±0.1°C specifically to suppress this.

3. Operator variability. In manual titration, different analysts judge the color endpoint differently. Studies of analytical chemistry reproducibility routinely find inter-analyst variability contributing 2–5% relative standard deviation to results — a Type A uncertainty source that disappears when automated spectrophotometric endpoints replace human judgment.

4. Sampling uncertainty. In environmental monitoring, the measurement of a contaminant in a water body is dominated not by instrument precision but by spatial heterogeneity. The uncertainty in the sampling strategy frequently exceeds the instrument uncertainty by an order of magnitude.

5. Model-dependent uncertainty. Radiocarbon dating (¹⁴C) requires a calibration curve — the IntCal20 dataset published by Radiocarbon — to convert measured isotope ratios to calendar years. Uncertainty in the calibration curve itself propagates into every date reported and can reach ±50 years or more for certain periods.

Decision boundaries

Uncertainty quantification becomes consequential when measurements drive decisions. Regulatory thresholds provide the clearest examples. The U.S. Environmental Protection Agency (EPA) sets maximum contaminant levels for drinking water at concentrations where measurement uncertainty relative to the threshold becomes a practical enforcement question — if a laboratory's expanded uncertainty for lead analysis is ±2 parts per billion and the action level is 15 parts per billion (EPA Lead and Copper Rule), a result of 14 ppb cannot be declared compliant or non-compliant without explicit uncertainty accounting.

The key decision heuristic is the guard band — a margin set inside a specification limit equal to the expanded uncertainty of the measurement system. Measurements outside the specification plus guard band fail unambiguously. Measurements inside the guard band pass unambiguously. The zone in between requires a risk-based judgment, often documented in a measurement uncertainty policy.

For scientific publishing, journals affiliated with the American Physical Society require expanded uncertainty to be reported for all key quantities, with coverage factor stated explicitly. Without that reporting, a result cannot be reproduced by an independent laboratory — and reproducibility sits at the center of the scientific enterprise as documented across this reference.