Measurement uncertainty in potency assayMeasurement Uncertainty (MU) applied to assay results

Measurement uncertainty is a parameter, associated with a measurement result, that characterizes the dispersion of values that could reasonably be attributed to the quantity being measured. It is not a mistake, not an error bar drawn as a courtesy, and not the same thing as a specification tolerance. Every potency result — 99.2%, 101.5%, 94.8% — carries an uncertainty interval whether or not the laboratory bothers to compute and report it; ISO/IEC 17025 clause 7.6 simply requires accredited labs to know what that interval is and to be able to state it.

The distinction that trips up non-metrologists: uncertainty is not the same as the acceptable range of individual results that a validated method's precision would predict (that's a repeatability statistic on its own), and it is not the specification itself (that's a product-quality limit set independently). Uncertainty describes how confident the lab can be that the reported single value is close to the item's true potency; the specification describes what value range is acceptable for that potency to be. The two must be reasoned about together at release, but they answer different questions.

The Guide to the Expression of Uncertainty in Measurement (GUM, published as JCGM 100:2008) is the internationally harmonized framework nearly all pharmaceutical, dietary supplement, and food testing labs use, directly or through a EURACHEM/CITAC adaptation, to build an uncertainty estimate. The GUM approach proceeds in defined steps:

1. Specify the measurand precisely — for a potency assay, this means stating the exact quantity being reported (e.g., % label claim of active X, on an as-is basis, by the validated HPLC method).
2. Identify every input quantity that contributes to the result — reference standard purity, sample and standard weighings, volumetric dilutions, chromatographic response/calibration, and analyst/instrument/day variability captured in repeatability and intermediate precision studies.
3. Quantify the standard uncertainty of each input — either by statistical analysis of repeated observations (Type A evaluation) or by other means such as certificate values, manufacturer specifications, or prior data (Type B evaluation).
4. Combine the standard uncertainties of all inputs into a combined standard uncertainty, propagating them according to the mathematical relationship connecting the inputs to the final result (the propagation-of-uncertainty law, accounting for correlations where they exist).
5. Apply a coverage factor (commonly k=2) to the combined standard uncertainty to obtain the expanded uncertainty, giving an interval with an approximate 95% level of confidence.

Some laboratories use a simplified 'top-down' approach instead of building a component-by-component budget from first principles — deriving the combined uncertainty directly from validation and proficiency-testing data (method precision, bias from recovery/reference-material studies, and reproducibility from interlaboratory or long-term QC data). Both bottom-up (component budget) and top-down (validation-data-derived) approaches are acceptable under ISO/IEC 17025, provided the method is documented and the result is defensible.

Once the individual standard uncertainties are estimated, they are combined (typically by root-sum-of-squares for independent, uncorrelated components) into the combined standard uncertainty u_c, and expanded uncertainty U = k × u_c, most commonly reported at k=2.

The combined standard uncertainty u_c represents roughly one standard deviation of the estimated distribution of possible true values. Reporting u_c alone at that coverage would mean the reported interval only has about a 68% chance of containing the true value — too narrow for most regulatory or release decisions. Multiplying by a coverage factor k=2 (assuming an approximately normal or Student's-t distribution with adequate effective degrees of freedom) produces an expanded uncertainty U with approximately 95% confidence of containing the true value — the convention nearly universally adopted in pharmaceutical and food testing reporting.

A result should be reported as, for example, '99.2% ± 1.4% (k=2),' making explicit both the coverage factor used and that the reported interval is the expanded, not standard, uncertainty. Reporting a bare '±' figure without stating the coverage factor is ambiguous and is itself flagged in ISO/IEC 17025 assessments.

Guard-banding narrows the effective acceptance range applied at release by the amount of the measurement uncertainty, so that a result is only accepted if it falls within the specification even after accounting for the possibility that the true value could be as far from the reported value as the uncertainty interval allows. For a specification of 95.0–105.0% and an expanded uncertainty of ±1.5%, a simple guard-banded acceptance zone becomes 96.5–103.5% — a result of 96.0% would be reported as passing on a bare point-comparison basis but rejected under the guard band because its uncertainty interval extends below the true specification limit.

ILAC-G8:09/2019 Guidelines on Decision Rules and Statements of Conformity is the framework accreditation bodies use to assess whether a laboratory's decision rule is fit for purpose. It formalizes the shared-risk versus guard-banded distinction above and requires the laboratory to: (1) determine, in consultation with the customer where relevant, an appropriate decision rule that reflects the level of risk acceptable for the application; (2) apply that rule consistently and record which rule was applied on the report or in a referenced document; and (3) understand that different customers or different products may justify different rules, provided each is documented and fixed in advance for that context.

In practice, a laboratory serving a pharmaceutical manufacturer under 21 CFR Part 211 will typically be expected, or contractually required, to apply a defined guard-banding approach for potency release testing given the higher consumer-risk consequence of releasing an out-of-specification drug product, while a lower-risk application (e.g., an internal in-process check with a downstream confirmatory test) may reasonably use simple shared-risk acceptance.

USP General Chapter <1010> Analytical Data — Interpretation and Treatment addresses how analytical results, including their variability, should be statistically evaluated and reported, and it underpins the expectation that a laboratory's OOS investigation (per FDA's 2022 revised OOS guidance) accounts for the assay's known measurement uncertainty rather than treating every point value as an exact truth.

An OOS investigation into a result that falls just outside specification should include, as part of Phase I laboratory investigation, a review of whether the result is consistent with the method's established measurement uncertainty and historical performance — not as grounds to invalidate the result, but as context that determines whether retesting, and how much of it, is scientifically justified. An OOT (out-of-trend) evaluation similarly benefits from a well-characterized uncertainty budget: a shift that is small relative to the assay's known intermediate precision may be indistinguishable from measurement noise, while a shift larger than the uncertainty budget would predict is a stronger signal of genuine process drift.

• State the expanded uncertainty and coverage factor explicitly wherever a conformity statement is made — '99.2% ± 1.4% (k=2)' rather than a bare percentage.
• Document, once, at the method or quality-system level, which decision rule (simple acceptance, guard-banded, or a defined partial guard band) applies for each type of conformity statement the lab issues, and apply it consistently.
• Re-evaluate the uncertainty budget when any major contributing element changes — a new reference standard lot with a wider certified uncertainty, a method revalidation, new instrumentation — rather than carrying forward a stale budget indefinitely.
• Distinguish, in every report and investigation, between measurement uncertainty (how confident we are in this result) and process/product variability (how much the true potency actually varies batch to batch) — conflating the two leads to either chasing measurement noise as a process problem or dismissing real drift as assay noise.
• When a customer or downstream process (e.g., a stability-indicating trend or a capability index) consumes the reported value, make clear whether that value already reflects a guard band or is the raw point estimate — mixing the two silently downstream produces double-counted or under-counted risk.