Accurate and auditable records of calibration-curve model selection for LC-MS assays are essential for reproducibility and transparent review. This article focuses on documenting the rationale and evidence used when choosing regression weighting, evaluating residuals, and disposing of back-calculated responses. It outlines practical record items and reviewer actions that preserve traceability without making performance claims.
Documenting weighting decisions
Weighting choices (for example, no weighting, 1/x, or 1/x2) should be recorded with a clear justification and supporting analytical evidence. Records should state the candidate models tested, the algorithm or software and version used to fit each model, and the numeric criteria used for selection (e.g., overall percent relative error, sum of squared residuals, or a formal heteroscedasticity test). Include the date and analyst who ran the comparisons, the dataset identifier, and any pre-specified acceptance rules. Cross-reference procedural standards such as regulatory and methodological guidelines when those standards are applicable; for procedural context see FDA bioanalytical method guidance and the ICH M10 guideline on bioanalytical method validation.
Residual diagnostics and fit assessment
Residual plots and summary statistics are primary analytical evidence for model suitability. Record the types of residuals plotted (absolute, percent, or studentized), axes and scaling used, and any automated or manual rules applied to interpret patterns (e.g., funnel shapes indicating heteroscedasticity). Document any outlier-detection methods and the timestamps and versions of files containing the plots. If a particular residual pattern led to selecting a different weighting or transformation, capture the decision trail: the observed pattern, the hypothesis for its origin, alternative models tested, and the final verdict with sign-off by the reviewer.
Back-calculated responses and reviewer disposition
Back-calculated concentrations are evidence of curve performance; maintain records of the full set of back-calculated values, the acceptance limits applied, and any points excluded from the final calibration. When exclusions occur, record the exclusion reason (e.g., preparation error, instrument alert), whether the point was re-assayed, and the impact on the model parameters and fit statistics. Reviewers should append a disposition note for each excluded or flagged response that explains the conditional next step (retain with justification, repeat analysis, or mark as not used for reporting). Include digital signatures or tracked reviewer identifiers and timestamps for these dispositions to support auditability.
Storage, versioning, and reproducibility
Store model-selection records in a manner that preserves raw data, intermediate files, and final reports. Recommended record elements include dataset IDs, software and script files (with checksums or version control links), a plain-language summary of the model-selection path, and reproducibility instructions sufficient for an independent analyst to rerun the selection. Link procedural references and any deviation logs to the calibration record. For statistical methods and assumptions used during selection, consider documenting references to standard statistical resources such as the NIST guidance on regression diagnostics where appropriate: NIST regression diagnostics.
Well-structured records enable reviewers to confirm that model-selection choices were data-driven and consistently applied. A concise record that ties weighting choice, residual evidence, and disposition of back-calculated responses to the final calibration allows traceable decisions and reduces ambiguity during subsequent reviews or audits.
For procedural alignment, reference the cited guidance documents and internal SOPs when drafting the record template and approval workflow. Ensure that the template prompts for the specific items described above to standardize entries across analysts and projects.
Not for human consumption. For laboratory research use only.