A calibration can look almost perfectly straight while still giving unsuitable errors at the concentrations that matter. Linearity describes a response relationship; a useful assay range also needs evidence that results are sufficiently accurate and precise throughout that interval. These questions should be assessed together without treating them as synonyms.
Separate the response model from the supported result interval
ICH Q2(R2) links the analytical range with suitable response, accuracy and precision between its lower and upper limits. It distinguishes the range of reportable results from the concentrations presented to the instrument through sample preparation.ICH — Q2(R2), range and response (opens in a new tab)
The practical question is not simply whether standards were prepared at several levels. It is whether the full procedure can produce suitable results at the levels intended for reporting.
| Range | What it describes |
|---|---|
| Prepared calibration levels | The standards included in a calibration exercise |
| Instrument working concentrations | The concentrations presented for measurement |
| Reportable material results | The original-sample values supported by the procedure |
A dilution can connect a high material concentration to a lower instrument concentration, but that preparation and its performance must be included in the supported procedure.
Inspect the differences left by the fitted model
NIST’s calibration guidance recommends examining residuals for structure that could invalidate the calibration relationship. A residual is the difference between an observation and the value predicted by the fitted model.NIST — Calibration data analysis and model validation (opens in a new tab)
A curved pattern of residuals can reveal a mismatch that is difficult to see on a graph spanning a large signal range. The overall correlation coefficient summarises association, not every local error.
For an original example, a calibration may span nominal concentrations from 1 to 1,000 units. A low standard back-calculated as 1.5 instead of 1.0 has a 50% error, even if the high-concentration points dominate the overall fit and make the plotted line look excellent.
The same principle applies at the upper end, where a response can depart from the chosen model. The instrument’s ability to display a signal does not establish valid quantification there.
Choose and assess the model rather than force a straight line
NIST describes considering a linear model and assessing whether higher-order terms are justified through the data and residual behaviour. A more complicated equation should solve a demonstrated modelling problem, not merely improve an attractive summary statistic.NIST — Calibration data analysis and model validation (opens in a new tab)
Weighting can also change which regions of a calibration have the greatest influence. Its justification depends on the error behaviour and analytical purpose; it is not a universal remedy chosen after seeing which model gives a desired sample result.
An independently prepared check sample can reveal a problem shared by all calibrators from one preparation. Recalculating the same standards through the same fitted model provides a narrower check.
Keep the model, weighting and acceptance rules with the data. If these change between runs, comparisons require evidence that the revised procedure remains suitable for the reported range.
Handle results outside the supported interval explicitly
A sample above the supported range may require an appropriate validated preparation or a bounded report. Extrapolating the line beyond its evidence can produce a precise-looking number without demonstrated performance.
At the lower end, distinguish a detectable signal from a reliably quantified result. A value below the quantification limit should not be promoted to an exact amount just because software calculates it.
For an original sample-basis example, an instrument working interval of 2–20 µg/mL corresponds to 20–200 µg/mL in an original liquid after a documented tenfold dilution. That arithmetic alone does not validate the dilution or extend the range further.
Sources and further detail
- ICH — Q2(R2), range and response (opens in a new tab)
Final November 2023 guideline, sections 2.3 and 3.2. Range is supported by response, accuracy and precision; no universal peptide range prescribed.
- NIST — Calibration data analysis and model validation (opens in a new tab)
Official statistical-methods handbook: model assessment and residual structure. Numerical examples are original and do not reproduce NIST data.
Sources checked 19 September 2026. Worked examples are illustrative unless a supplied report is explicitly identified. This article has not undergone independent scientific peer review.