Treatment of an observed bias

Importance of the topic

The control and treatment of observed bias in analytical measurements is central to producing reliable, traceable results and defensible measurement uncertainty (MU). Systematic effects influence decision-making in compliance testing, quality assurance and research. Clear rules for when to eliminate, correct or account for bias in the uncertainty budget ensure consistent reporting, comparability between laboratories and compliance with international guidance such as the GUM.

Objectives and overview of the leaflet

This leaflet explains the principles for deciding whether to correct an observed significant bias and how such a decision impacts measurement uncertainty. It outlines key criteria for correction, a practical decision roadmap, and options when correction is not applied. The leaflet is aimed at laboratories validating standardised or in‑house analytical methods and covers considerations for both method and laboratory bias.

Methodology and decision criteria

The leaflet bases decisions on the following essential questions:

• Is the cause of the observed bias understood?
• Can the magnitude of the bias be determined reliably?
• Is the bias consistent across all test samples within the method scope?
• Should any correction be multiplicative or additive (i.e., does the bias vary with concentration)?

Key procedural steps and principles:

1. Assess the size of the observed bias—small biases that are not practically important rarely justify the resource cost of elimination or correction.
2. If bias is non‑negligible, prioritise elimination through method modification where feasible (e.g., reduce interferences, improve calibration).
3. If elimination is impractical, consider correction only after determining whether correction is required by regulation, forbidden, or allowed.
4. When correction is allowed, apply it only if the bias can be reliably estimated and the correction method is applicable across the method scope; unknown causes or unreliable estimates advise against correction.
5. Implement corrections only if they yield a meaningful reduction in MU: the uncertainty introduced by the correction must be smaller than the component of MU that would remain uncorrected.

The leaflet emphasises that for empirical (operationally defined) methods, method bias is defined as zero, but laboratory bias (e.g., due to procedure, operator or equipment) must still be assessed.

Main findings and discussion

Principal conclusions drawn in the leaflet include:

• GUM expects results to be corrected for recognised significant systematic effects; therefore, investigation and minimisation of known biases during method development is expected.
• If a significant bias remains after development, validation must quantify any additional bias across relevant concentrations and matrices.
• Correcting for bias without a reliable bias estimate can increase rather than decrease MU and is therefore discouraged.
• If correction is impractical or not permitted, several pragmatic options exist (not exhaustive):

Options when no correction is applied (illustrated specifically for recovery bias but applicable to other biases):

• Do nothing if the bias is negligible for the intended use.
• Report the uncorrected result together with recovery information and the uncertainties for both result and recovery.
• Incorporate the bias into the uncertainty estimate of reported results.

The leaflet also references literature that surveys additional strategies and modelling approaches for accounting for bias in MU estimates.

Contributions and practical implications

The leaflet provides a concise, practical framework laboratories can adopt during method validation and routine analysis. It clarifies when resources should be devoted to eliminating bias versus applying corrections, and it stresses that any correction must improve overall measurement quality by reducing combined MU. For routine practice this translates into:

• Structured investigations of bias sources during method development and validation.
• Documented criteria for when correction is applied, including uncertainty evaluation for the correction itself.
• Transparent reporting strategies when correction is not made, aiding comparability and decision-making by end users.

Future trends and potential applications

Expected developments and useful directions include:

• Improved statistical and computational tools for robustly estimating bias and propagating uncertainty, including Bayesian and Monte Carlo approaches.
• Standardised protocols and acceptance criteria for bias assessment across matrices and concentration ranges, improving inter‑laboratory comparability.
• Increased use of proficiency testing and reference materials to distinguish laboratory bias from method bias and to validate correction procedures.
• Automation of bias-tracking within laboratory information management systems (LIMS) to flag persistent biases and document corrective actions.
• Further guidance on choosing multiplicative versus additive corrections and on combining correction uncertainty with other MU components.

Conclusion

Treatment of an observed significant bias should follow a reasoned, documented decision process: attempt elimination first; if impractical, consider correction only when the cause and magnitude of bias are well understood and the correction reduces overall measurement uncertainty. When correction is not applied, laboratories must adopt clear reporting or uncertainty-adjustment strategies to ensure results remain useful and comparable. Adhering to these principles aligns practice with the GUM expectation to correct recognised systematic effects where appropriate.

Reference

1. JCGM 100:2008. Evaluation of measurement data – Guide to the expression of uncertainty in measurement (GUM).
2. Harmonised guidelines for the use of recovery information in analytical measurement. Pure and Applied Chemistry, Vol. 71, No. 2, pp. 337–348, 1999.
3. Magnusson, B.; Ellison, S. L. R. Analytica Chimica Acta / Analytical and Bioanalytical Chemistry review-style coverage on bias and uncertainty, 2008, 390, 201–213.