The Gold Standard of Scientific Instrumentation | Same Reading, Different Measurement Uncertainty

Two instruments can report the same corrected 25.0 °C with very different standard and expanded uncertainties. These original budgets make the difference auditable.

Chapter index

Advanced companion: uncertainty and guard bands

Chapters 1–4

1. An identical display is not identical measurement evidence · 2. Define the measurand, result and units · 3. Distinguish repeatability from calibration information · 4. Construct Instrument A’s budget

Chapters 5–8

5. Construct Instrument B’s larger budget · 6. Expanded uncertainty requires a declared coverage factor · 7. Calculate and label the illustrative intervals · 8. Introduce an illustrative acceptance threshold carefully

Chapters 9–12

9. Why the coverage interval is not the whole decision rule · 10. Distinguish a known systematic correction from uncertainty · 11. Avoid double-counting or ignoring correlations · 12. Resolution and the number of displayed digits

Chapters 13–16

13. Repeatability of a mean depends on sample size · 14. Traceability is not the same as small uncertainty · 15. Frequently asked questions and changed assumptions · 16. Apply the Gold Standard of measurement evidence

Two hypothetical temperature instruments each show 25.0 degrees Celsius after any required known corrections. If a dashboard displays just these values, a reader may conclude that the instruments agree completely and support the same confidence in the underlying temperature.

The displayed central values do agree. But they may carry different uncertainties because of repeatability, calibration evidence, measurement conditions and the way the result was obtained. The same numerical reading does not determine how precisely the measurand has been estimated.

We construct Instrument A with two independent standard uncertainty components of 0.10 °C and 0.20 °C. Instrument B has corresponding components 0.40 °C and 0.60 °C. The first component in each case describes the uncertainty associated with the reported mean; the second represents calibration-related standard uncertainty.

These are invented instrument records for mathematical teaching, not actual certificates or a claim about named thermometers. We will combine the components using the declared independence model and examine what the different resulting intervals mean for an illustrative threshold.

Contents · Next chapter

The measurand is the temperature of a specified location or material under a defined measurement procedure. A reported estimate of 25.0 °C needs context: what was measured, when, where and under which environmental conditions?

A measurement result is more than the numerical estimate. It also needs an indication of measurement uncertainty and information about how the result relates to reference standards and the measurement model.

In our example the numerical result y=25.0 °C is the same for both instruments. Their uncertainty contributions are expressed in Celsius degrees, equivalent in size to kelvin increments. We are dealing with differences and uncertainty widths, not raising absolute temperatures to powers.

No real specimen acceptance decision can be made from this invented result alone. The purpose is to calculate how different uncertainty budgets attach different evidential meaning to an otherwise identical displayed value.

Contents · Previous chapter · Next chapter

A Type A evaluation of uncertainty uses statistical analysis of observed data, such as repeated readings under specified conditions. If a reported result is the mean of n independent repeat measurements with sample standard deviation s, a conventional standard uncertainty of that mean is s/√n.

A Type B evaluation uses other information, which may include calibration certificates, reference standards, specifications and prior characterisation. Type B does not automatically mean guesswork; it means evaluation by methods other than a direct statistical analysis of the current repeated observations.

Our two uncertainty components are already expressed as standard uncertainties for the reported measurand. We do not treat a certificate’s expanded uncertainty as a standard deviation without accounting for its coverage factor.

These distinctions follow the approach in NIST Technical Note 1297 on combined standard uncertainty. An uncertainty budget should name each contributor, its statistical or documentary basis and its relationship to the final result.

Contents · Previous chapter · Next chapter

For Instrument A, take a repeatability-related standard uncertainty u_r=0.10 °C for the reported mean and a calibration-related standard uncertainty u_c=0.20 °C. Assume their effects are independent under the chosen measurement model.

The combined standard uncertainty is u_A=sqrt(u_r²+u_c²)=sqrt(0.10²+0.20²)=sqrt(0.05)≈0.223607 °C. It is not the ordinary sum 0.30 °C, because independent standard uncertainties combine by root-sum-of-squares in this linear model.

The result can therefore be reported as an estimate y_A=25.0 °C with combined standard uncertainty about 0.224 °C, accompanied by the assumptions and input sources.

This is not proof that the actual temperature differs from 25.0 by less than 0.224. A standard uncertainty represents an estimated dispersion scale associated with the measurement result, not a deterministic error bound.

Contents · Previous chapter · Next chapter

Instrument B shows the same corrected numerical estimate of 25.0 °C. Its stipulated repeatability-related standard uncertainty is 0.40 °C and its calibration-related standard uncertainty is 0.60 °C.

Under the same independence assumption, combined standard uncertainty is u_B=sqrt(0.40²+0.60²)=sqrt(0.16+0.36)=sqrt(0.52)≈0.721110 °C.

Both instruments therefore have the same displayed central estimate but different combined standard uncertainties: about 0.224 °C versus 0.721 °C. Instrument B’s uncertainty scale is roughly 3.23 times larger.

Nothing in the displayed number of decimal places establishes this distinction. Two instruments can both show one decimal place, while their calibration and measurement-process uncertainty budgets differ materially.

Contents · Previous chapter · Next chapter

In some reporting settings, an expanded uncertainty U=k×u_c is reported rather than only the combined standard uncertainty. The coverage factor k is chosen to support an interval under a specified statistical interpretation and uncertainty model.

For a purely illustrative choice k=2, Instrument A has expanded uncertainty U_A≈2×0.223607=0.447214 °C, conventionally rounded here to 0.45 °C. Instrument B has U_B≈2×0.721110=1.442221 °C, rounded to 1.44 °C.

The NIST guidance on expanded uncertainty and coverage factors explains that k=2 often approximates 95-percent coverage under suitable normal-distribution and reliable-standard-uncertainty assumptions. It is not an unconditional guarantee.

The appropriate reporting precision and coverage factor depend on the measurement, degrees of freedom, distribution and intended use. This example uses one simple k for an arithmetic comparison, not a substitute for a formal uncertainty evaluation.

Contents · Previous chapter · Next chapter

Using the rounded expanded uncertainties, Instrument A’s interval centred on 25.0 °C is approximately 24.55 to 25.45 °C. Instrument B’s is approximately 23.56 to 26.44 °C.

These are intervals describing the reported expanded uncertainty under the stated model and rounding. They should not be described as absolute physical limits that the true temperature can never cross.

The large difference in interval width demonstrates why one common reading does not give two instruments equal decision power. Instrument A constrains the temperature more tightly under its uncertainty assumptions.

The main lesson is about documentation: a dashboard should not conceal uncertainty and then assign the same confidence to two values merely because their displayed central estimates and decimal formats agree.

Evidence quantity	Instrument A	Instrument B
Corrected reading	25.0 °C	25.0 °C
Repeatability standard u	0.10 °C	0.40 °C
Calibration standard u	0.20 °C	0.60 °C
Combined standard u	0.2236 °C	0.7211 °C
Expanded U with k=2	0.4472 °C	1.4422 °C
Illustrative interval (rounded)	24.55–25.45 °C	23.56–26.44 °C

Contents · Previous chapter · Next chapter

Suppose a fictional non-safety-critical acceptance exercise considers a temperature upper limit of 25.8 °C. This is an invented threshold used only to demonstrate a guard-banded decision rule, not a real legal, clinical or engineering specification.

Choose the explicit rule: accept only if y+U≤25.8 °C using our reported expanded uncertainty U. For Instrument A, y+U≈25.45 °C, so the numerical rule permits acceptance. For Instrument B, y+U≈26.44 °C, so the rule does not permit acceptance.

The second outcome is not proof that the actual temperature exceeds 25.8 °C. It means the selected conservative acceptance rule lacks sufficient evidence for acceptance based on Instrument B’s uncertainty budget.

Real conformity assessment rules depend on permissible false acceptance and false rejection risks, relevant standards and consequences. One should not adopt this invented guard band for actual equipment without the appropriate decision framework.

Contents · Previous chapter · Next chapter

An expanded uncertainty interval expresses a distribution-related range around a measured estimate under assumptions. Whether that range must lie entirely below a limit depends on the chosen conformity-assessment decision rule.

Some applications use guard bands, shared risk, conditional acceptance or specified consumer and producer risks. Others require repeated measurements or different levels of coverage. Reporting y±U alone does not choose among these policies.

In our teaching example the same corrected y appears twice, but one uncertainty interval crosses the fictional decision limit. The different outcomes are a consequence of the explicit decision rule plus uncertainty information, not a moral ranking of the two instruments.

A credible report identifies the measurement result, uncertainty definition, acceptance limit and decision rule separately. A reader should be able to reproduce the decision without guessing which convention was used.

Contents · Previous chapter · Next chapter

Suppose a calibration exercise identifies a known offset that must be corrected. A good measurement procedure applies the correction to the raw indication when appropriate and retains uncertainty about the correction in its budget.

It would be incorrect to knowingly leave a significant identified bias uncorrected and treat the combined standard uncertainty as if it automatically removed that bias. An uncertainty statement is not a substitute for applying recognised systematic corrections.

Our instruments are stated to report already corrected estimates of 25.0 °C, so the example is about residual uncertainty after known corrections rather than hidden biased central values.

The NIST combined-uncertainty guidance notes the expectation that corrections for recognised systematic effects are applied where appropriate. The difference between correction and uncertainty is fundamental to traceable measurement.

Contents · Previous chapter · Next chapter

Root-sum-of-squares combining is valid for the simple two-component model when the input uncertainties are appropriately treated as independent and sensitivity coefficients are one. It is not a universal formula that allows correlated effects to be ignored.

If calibration components share the same reference standard or environmental drift affects several readings together, covariance terms may be required. Correlation can increase or decrease combined uncertainty depending on the model and direction of the common effect.

Repeated measurements can reduce some random uncertainty of the mean as n grows, but collecting more values does not automatically eliminate a shared calibration offset. Treating all sources of uncertainty as independent repeatability noise would overstate the benefit of repetition.

An uncertainty budget should document the dependency structure and explain how components were evaluated. More mathematical detail is justified when the decision is sensitive to those assumptions.

Contents · Previous chapter · Next chapter

A display showing 25.0 °C has resolution of 0.1 °C under a simple rounded display model. That does not mean its combined uncertainty is automatically ±0.05 °C or that its accuracy is assured to a tenth of a degree.

Digitisation and rounding can contribute to an uncertainty budget. If a display rounds uniformly within a width of 0.1 °C and assumptions justify a rectangular distribution, the associated standard uncertainty is 0.1/√12≈0.0289 °C.

For simplicity our initial two-component budgets stipulate that any relevant resolution contribution has been adequately included or is negligible for the illustrated comparison. An actual evaluation should make that treatment explicit.

The number of digits shown by a display, the statistical repeatability of readings and the calibration traceability of the instrument are distinct properties. A clean readout alone is not enough evidence for a small measurement uncertainty.

Contents · Previous chapter · Next chapter

If twenty-five independent replicate readings have sample standard deviation 0.5 °C under stable conditions, the conventional Type A standard uncertainty of their mean is 0.5/√25=0.10 °C.

This provides one possible mathematical origin for Instrument A’s stipulated 0.10 °C repeatability-related component. Instrument B could have a different replicate variance or measurement design that yields the larger 0.40 °C contribution.

The 1/√n reduction requires conditions such as independence and a stable measurand. Strong serial correlation or systematic drift can make extra repeated readings less effective than the naive formula predicts.

For a real instrument, state whether the reported result is a single indication, an average over repeated readings or the output of a filtering algorithm. The standard uncertainty must correspond to the actual reported quantity.

Contents · Previous chapter · Next chapter

Metrological traceability concerns a documented chain relating a measurement result to appropriate references, typically through calibrations each contributing uncertainty. Traceability does not imply zero uncertainty or that two traceable devices will have equally narrow intervals.

Calibration documentation should identify standards, environmental conditions, calibration range, corrections and uncertainty contributions. A traceable reading can still be unsuitable for a narrow tolerance when its uncertainty is too large.

Conversely, a device with impressive short-term repeatability may lack a defensible relationship to the intended temperature scale if its calibration evidence is missing. Precision and traceability answer different questions.

The practical evaluation combines calibration evidence, repeatability, installation conditions and the measurand definition instead of treating one pleasant-looking display as a complete measurement report.

Contents · Previous chapter · Next chapter

Does 25.0 °C on both devices prove they agree exactly? It shows their displayed central estimates agree. It does not prove the true temperature or measurement errors are identical.

Can k=2 always be labelled exactly 95 percent? No. That approximate interpretation depends on the uncertainty model and degrees of freedom. The factor must be declared and justified.

What if the repeatability component becomes half as large? Recalculate root-sum-of-squares while keeping calibration uncertainty unchanged. The combined improvement may be smaller than expected if calibration dominates the budget.

Practice: for a hypothetical instrument with independent standard uncertainty components 0.30 °C and 0.40 °C, calculate combined standard uncertainty 0.50 °C and k=2 expanded uncertainty 1.00 °C. Explain why simply adding 0.30 and 0.40 would not represent the same uncertainty calculation.

Contents · Previous chapter · Next chapter

Our two fictional instruments each report corrected 25.0 °C, yet the combined standard uncertainties are about 0.224 °C and 0.721 °C. Under illustrative k=2 reporting, their expanded uncertainties are about ±0.45 °C and ±1.44 °C.

Continue with The Gold Standard of Scientific Instrumentation and The Gold Standard of Uncertainty Quantification for related measurement and modelling principles.

Primary references are NIST Technical Note 1297 on Combined Standard Uncertainty and NIST Expanded Uncertainty and Coverage Factors. They supply the technical concepts; both instrument budgets, temperature readings and the decision threshold are original hypothetical examples.

Use the Gold Standard hub for connected topics. A dependable numerical reading is accompanied by its measurand definition, calibration status, uncertainty method and decision rule, rather than by a decimal display alone.

Contents · Previous chapter

Advanced worked companion: traceability, covariance and sensitivity

Reconstruct the calibration component from a certificate

A calibration certificate may report an expanded uncertainty U rather than a standard uncertainty. If a certificate states U_cal=0.40 °C with coverage factor k=2 for the relevant calibrated quantity, the corresponding standard uncertainty is U_cal/k=0.20 °C under the certificate’s stated interpretation. That could supply the calibration component for Instrument A’s simple budget.

Likewise, a certificate stating an expanded calibration uncertainty of 1.20 °C with k=2 could imply a calibration-related standard contribution of 0.60 °C for Instrument B, if the calibration model and quantity are directly applicable. The published factor and confidence interpretation matter; a number labelled U must not be inserted as though it were already a standard deviation.

The certificate also needs to cover the relevant temperature range, measurement configuration and calibration status. An uncertainty figure obtained at one operating point cannot always be transferred unchanged to a different temperature, immersion depth or environmental condition. Drift since calibration may require an additional component.

A complete budget records the documentary source of each Type B contribution, the applied corrections and whether the correction has been used properly. Metrological traceability identifies a chain of comparisons and uncertainty contributions; it does not mean the chain is error-free or that the final uncertainty is automatically small.

Derive a Type A uncertainty of a measured mean

Suppose a series of twenty-five statistically independent readings has sample standard deviation s=0.5 °C and the measurand is estimated by their arithmetic mean. Under the conventional repeatability model, the standard uncertainty of that mean is s/√n=0.5/√25=0.10 °C. This matches the invented repeatability component for Instrument A.

For another series of twenty-five independent readings with s=2.0 °C, the same calculation yields 0.40 °C for the reported mean. This would be one plausible mathematical explanation for Instrument B’s larger repeatability component. The two example standard deviations do not establish anything about actual thermometer performance.

The standard deviation of individual indications differs from the standard uncertainty of their estimated mean. Using s rather than s/√n for the latter can overstate that component if the independence and fixed-measurand assumptions hold. Conversely, using s/√n in a strongly drifting or serially correlated sequence can understate uncertainty.

An adequate evaluation records sample size, the observations, measurement stability and whether repeated measurements share environmental or calibration influences. Gathering many repeated values can reduce random noise without eliminating an uncertainty common to every reading.

What happens if uncertainty contributors are correlated?

For two components with standard uncertainties u₁ and u₂ and correlation coefficient ρ, a linear combination with unit sensitivity coefficients can have combined variance u_c²=u₁²+u₂²+2ρu₁u₂. The familiar root-sum-of-squares expression assumes the cross term is zero.

For Instrument A’s component magnitudes 0.10 and 0.20 °C, independence gives u_c=sqrt(0.05)≈0.223607 °C. If a substantiated correlation of ρ=0.5 applied, the combined variance would be 0.01+0.04+0.02=0.07, producing u_c≈0.264575 °C and a k=2 expanded uncertainty about 0.52915 °C.

For a correlation of ρ=−0.5, the same arithmetic gives combined variance 0.03 and u_c≈0.173205 °C. These scenarios are hypothetical sensitivity checks. One must not choose a convenient negative correlation merely to report a narrower interval; the dependence must arise from evidence about the measurement model.

Shared references, common environmental drift and repeated reuse of an uncertain calibration correction can create correlations. A transparent uncertainty budget states the assumptions and investigates whether the final result is sensitive to reasonable alternatives.

Add display resolution explicitly rather than guessing from digits

If a digital instrument rounds to the nearest tenth of a degree, a simple model places its rounding deviation within approximately ±0.05 °C. Under a rectangular distribution, the associated standard uncertainty is 0.05/√3=0.1/√12≈0.028868 °C.

If this were an additional independent contributor not already included in Instrument A’s other components, the revised combined standard uncertainty would be sqrt(0.10²+0.20²+0.1²/12)≈0.225462 °C. Its illustrative k=2 expanded uncertainty would be approximately 0.450925 °C.

That modest change occurs because rounding uncertainty is small relative to the assumed calibration term. For a higher-precision instrument, resolution could play a larger relative role. The important lesson is to document whether a component has been included, not to automatically add it twice.

The display’s number of digits is not a general accuracy guarantee. A reading can be repeatable and finely resolved while biased, drifting or improperly calibrated. Likewise, a coarser display may be sufficient for a task whose tolerance is wide relative to the combined uncertainty.

Decision rules should state the risk being controlled

Return to the fictional upper limit 25.8 °C. Under the explicitly chosen acceptance rule y+U≤25.8, Instrument A at 25.0 with U≈0.45 satisfies the condition because 25.45≤25.8. Instrument B at 25.0 with U≈1.44 does not satisfy it because 26.44>25.8.

This is a classification under one decision rule, not a probability that a physical sample passes or fails. In a formal conformity assessment, false-accept and false-reject risks, the uncertainty distribution and any applicable standard must guide selection of the rule.

The analyst should report whether the interval means expanded measurement uncertainty about the corrected result, whether k is justified, and how the acceptance limit is defined. A result just outside a guard band need not imply that the measurand is definitely nonconforming.

If a decision is safety critical or legally regulated, the appropriate process requires qualified metrology and the current relevant requirements. The invented 25.8 °C threshold is deliberately not tied to any real product standard.

A defensible instrument comparison

A complete measurement comparison identifies the measurand and sampling conditions, corrected central estimate, Type A sources, calibration and other Type B sources, covariance treatment, coverage factor and final expanded uncertainty. It also describes the traceability evidence and any limitations of the calibration range.

For Instrument A and B, the central readings are identical but the standard uncertainties are not. This does not alone prove Instrument A is a better product in every dimension; speed, operating range, environmental robustness and cost can also matter. It establishes a narrower reported uncertainty under the chosen model.

Measurement uncertainty is a property of the result and its method, not merely a fixed adjective attached to a piece of hardware. Changing environment, averaging time or calibration status can change the uncertainty budget, even when the displayed temperature remains 25.0 °C.

The relevant foundations are NIST TN 1297 on combined uncertainty and NIST guidance on expanded uncertainty. Both instruments, budgets and the threshold are original hypothetical constructions.

Return to chapter index · Scientific Instrumentation subject root