Total error or measurement uncertainty? Both combine bias and imprecision into a single figure of merit. The laboratory community has argued about which to use for years. They are not the same quantity. They answer different questions. The disagreement is really about what a result’s error is.

Total analytical error (TAE) and measurement uncertainty (MU) both combine bias and imprecision into one number that summarises how wrong a result might be. Both use the same two ingredients, so they are often treated as interchangeable. They are not. The two combine those ingredients differently, and they describe different things. Choosing between them is really a choice about how you conceive of a result’s error.

Total error: the worst plausible single result

The total analytical error model asks a blunt, practical question: how far off could a single result be, at worst? The model adds bias to a multiple of the imprecision, so that bias plus 1.65 or 1.96 standard deviations bounds the error of an individual measurement. Bias and imprecision are treated as acting in the same direction, because for one result on one day they might. The output is a one-sided worst-case envelope. Compare it against an allowable total error to decide whether the method is fit for a clinical decision.

With a bias of 1.0% and an imprecision of 2.0%, the total error at 1.65 SD is 1.0% + 1.65 × 2.0% = 4.3%. That 4.3% is the figure to set against the allowable limit. CLSI EP21 formalises estimating total error from a comparison study.

Measurement uncertainty: the dispersion of the value

Measurement uncertainty asks a different question: given this reported value, over what interval could the true value reasonably lie? The framework assumes known biases have been corrected, then characterises the remaining dispersion. The uncertainty of that correction combines with the imprecision in quadrature, as independent variances, to give a symmetric ±U interval around the result. Measurement uncertainty is therefore a statement about the value’s spread, expressed with a stated coverage, and it is what ISO 15189 and the metrological tradition ask for.

Two representations of a result's error: a total-error model adding bias and a multiple of the SD into a one-sided worst-case envelope compared against an allowable total error line; and a measurement-uncertainty model showing a symmetric interval around a bias-corrected value combining components in quadrature.
Total error adds bias to a multiple of the imprecision for a worst-case bound on a single result. Measurement uncertainty corrects the bias and combines the remaining components in quadrature into a symmetric interval around the value.

Why the difference matters

The two can give materially different figures from the same validation data. Each is right for its own purpose.

Where a method must be judged fit to support a clinical decision limit — will a single result mislead a treatment choice? — the worst-case total-error view against an allowable limit is the natural test, and it underlies method-decision approaches to setting quality goals.

Where a result must be reported with an interval that expresses how well the value is known, in the metrological sense an accreditation body expects, measurement uncertainty is the right instrument. Treating one as a direct substitute for the other is where the confusion, and the arguments, come from.

Which to report

In practice, laboratories often need both, for different audiences. The pragmatic position: use total error against an allowable limit to decide whether a method is good enough for its clinical purpose. Use measurement uncertainty to state, on a report or under accreditation, how well a given result is known. Whichever you quote, name it explicitly. A single “±” figure with no framework attached is ambiguous. The reader cannot tell whether bias was added or corrected, or whether the interval is worst-case or dispersion.

Common mistakes

Treating TAE and MU as the same number. One adds bias, the other corrects it; one is worst-case, the other is dispersion. They will differ, and both can be correct.

Adding components when you should combine in quadrature, or the reverse. The total-error model adds; the uncertainty model takes the square root of the sum of squares. The method must match the framework.

Reporting a bare ± figure. Without naming the framework and the coverage, the interval cannot be interpreted. State which you used.

Using a worst-case bound where dispersion is asked for. An accreditation body asking for measurement uncertainty is not asking for allowable total error. Give the quantity requested.

Work in either framework with Analyse-it

Analyse-it produces what both frameworks need from the same studies, inside Excel:

  • Total analytical error, combining bias from a method comparison with imprecision, assessed against an allowable total error (EP21-A)
  • The precision and bias components a measurement uncertainty is assembled from
  • So the choice of framework does not decide which studies you have to run

Every feature from all five editions for 15 days, with no sign-up and no licence key. Total error and the underlying precision and bias estimation are in the Method Validation and Ultimate editions, from US$ 475 a year. Validated against NIST and CLSI reference datasets. See total analytical error explained and estimating measurement uncertainty top-down.