6SQ ToolsWhat Is Calibration Curve Linearity and How Do You Evaluate It (ISO 8466-1)?
In analytical chemistry and quality control, a calibration curve links an instrument's response to known analyte concentrations. But a curve is only useful if it is linear — meaning the response changes proportionally with concentration across the working range. Evaluating calibration curve linearity is a core requirement in method validation, and ISO 8466-1 and ICH Q2 provide the authoritative framework for doing so.
---
What It Is
Calibration curve linearity is the ability of an analytical method to produce results directly proportional to the concentration of the analyte within a given range. It is not just about drawing a straight line through points — it is a statistical assessment of whether the linear model adequately describes the data, and whether deviations are small enough to be ignored.
ISO 8466-1 defines a standardized procedure for constructing and validating a linear calibration function, including statistical tests for outliers, homogeneity of variance, and lack-of-fit. ICH Q2 (Validation of Analytical Procedures) similarly requires linearity to be demonstrated, typically by visual inspection of the plot, the correlation coefficient (r), the coefficient of determination (R²), and the y-intercept.
---
How It Works: The ISO 8466-1 Approach
ISO 8466-1 gives a step-by-step statistical procedure. Here is the essence:
### 1. Experimental design
### 2. Fit the linear model
Fit the ordinary least-squares regression:
\[
y = a + b \cdot x
\]
where:
### 3. Check homogeneity of variance
ISO 8466-1 requires that the variance of the responses does not change systematically with concentration. This is checked by comparing the variance at the lowest and highest concentration levels using an F-test (at a chosen significance level, typically 95% or 99%).
### 4. Test for outliers
Apply a Grubbs test (or similar) to the residuals or to the replicate means to detect any single outlying point. If an outlier is found, investigate the cause; do not remove data arbitrarily.
### 5. Test for lack-of-fit
This is the critical step. The residual variance from the regression is compared with the "pure error" variance obtained from replicates at the same concentration level. The test statistic is:
\[
F = \frac{MS_{\text{lack-of-fit}}}{MS_{\text{pure error}}}
\]
If the calculated F is less than the critical F-value (with appropriate degrees of freedom), the linear model is adequate — i.e., there is no significant lack-of-fit.
### 6. Report the key statistics
Classic acceptance values: For ICH Q2, \(R^2 \geq 0.99\) (or \(r \geq 0.995\)) is commonly cited as a guideline, but ISO 8466-1 does not rely on R² alone — it emphasizes the statistical tests above. A high R² does not guarantee linearity if the data show systematic curvature.
---
A Worked Illustrative Example
Suppose you measure an absorbance response at six concentration levels (0, 1, 2, 3, 4, 5 mg/L), with three replicates each:
Concentration (mg/L) | Replicate responses (absorbance)
Step 1 — Regression: Fitting \(y = a + b x\) gives approximately:
\[
y = 0.0015 + 0.1503 \, x
\]
with \(R^2 = 0.9997\).
Step 2 — Variance check: Variance at 0 mg/L ≈ \(1.0 \times 10^{-6}\); variance at 5 mg/L ≈ \(1.2 \times 10^{-5}\). The F-ratio is about 12, which may exceed the critical F at 95% (with 2 and 2 degrees of freedom, \(F_{\text{crit}} \approx 19\)). Here it passes, but this illustrates why the check matters.
Step 3 — Lack-of-fit test: The pure error variance from replicates is small. The lack-of-fit mean square is computed from the deviation of each level's mean from the regression line. The resulting F is, say, 1.8, well below the critical value of ~3.7 (for 4 and 12 degrees of freedom at 95%). Conclusion: No significant lack-of-fit — the linear model is adequate.
Step 4 — Report: Slope = 0.1503, intercept = 0.0015, \(R^2 = 0.9997\), lack-of-fit F = 1.8 (not significant). The calibration curve is linear over 0–5 mg/L.
---
Common Pitfalls
---
Closing
Calibration curve linearity is not a single number — it is a statistical verdict based on variance checks, outlier tests, and lack-of-fit analysis, as prescribed by ISO 8466-1 and ICH Q2. To apply this method quickly and correctly, try the free Calibration Curve Linearity calculator at https://www.6sq.com/tools/calcurve/ — it guides you through the ISO 8466-1 steps and gives you the statistical verdict in seconds.
---
What It Is
Calibration curve linearity is the ability of an analytical method to produce results directly proportional to the concentration of the analyte within a given range. It is not just about drawing a straight line through points — it is a statistical assessment of whether the linear model adequately describes the data, and whether deviations are small enough to be ignored.
ISO 8466-1 defines a standardized procedure for constructing and validating a linear calibration function, including statistical tests for outliers, homogeneity of variance, and lack-of-fit. ICH Q2 (Validation of Analytical Procedures) similarly requires linearity to be demonstrated, typically by visual inspection of the plot, the correlation coefficient (r), the coefficient of determination (R²), and the y-intercept.
---
How It Works: The ISO 8466-1 Approach
ISO 8466-1 gives a step-by-step statistical procedure. Here is the essence:
### 1. Experimental design
- Prepare at least 5–6 concentration levels (ISO 8466-1 recommends 6), evenly spaced across the working range.
- Measure each level in duplicate or triplicate (ISO 8466-1 uses a specific replicate scheme; at minimum, independent measurements per level).
### 2. Fit the linear model
Fit the ordinary least-squares regression:
\[
y = a + b \cdot x
\]
where:
- \(y\) = measured response (e.g., absorbance, peak area)
- \(x\) = concentration
- \(b\) = slope (sensitivity)
- \(a\) = intercept
### 3. Check homogeneity of variance
ISO 8466-1 requires that the variance of the responses does not change systematically with concentration. This is checked by comparing the variance at the lowest and highest concentration levels using an F-test (at a chosen significance level, typically 95% or 99%).
### 4. Test for outliers
Apply a Grubbs test (or similar) to the residuals or to the replicate means to detect any single outlying point. If an outlier is found, investigate the cause; do not remove data arbitrarily.
### 5. Test for lack-of-fit
This is the critical step. The residual variance from the regression is compared with the "pure error" variance obtained from replicates at the same concentration level. The test statistic is:
\[
F = \frac{MS_{\text{lack-of-fit}}}{MS_{\text{pure error}}}
\]
If the calculated F is less than the critical F-value (with appropriate degrees of freedom), the linear model is adequate — i.e., there is no significant lack-of-fit.
### 6. Report the key statistics
- Slope \(b\) and intercept \(a\)
- Residual standard deviation \(s_{y,x}\)
- Correlation coefficient \(r\) and/or \(R^2\)
- Result of the lack-of-fit F-test
Classic acceptance values: For ICH Q2, \(R^2 \geq 0.99\) (or \(r \geq 0.995\)) is commonly cited as a guideline, but ISO 8466-1 does not rely on R² alone — it emphasizes the statistical tests above. A high R² does not guarantee linearity if the data show systematic curvature.
---
A Worked Illustrative Example
Example data (illustrative only) — not from a real laboratory.
Suppose you measure an absorbance response at six concentration levels (0, 1, 2, 3, 4, 5 mg/L), with three replicates each:
Concentration (mg/L) | Replicate responses (absorbance)
- 0 | 0.002, 0.001, 0.003
- 1 | 0.151, 0.149, 0.152
- 2 | 0.301, 0.298, 0.305
- 3 | 0.452, 0.448, 0.455
- 4 | 0.598, 0.602, 0.605
- 5 | 0.752, 0.748, 0.755
Step 1 — Regression: Fitting \(y = a + b x\) gives approximately:
\[
y = 0.0015 + 0.1503 \, x
\]
with \(R^2 = 0.9997\).
Step 2 — Variance check: Variance at 0 mg/L ≈ \(1.0 \times 10^{-6}\); variance at 5 mg/L ≈ \(1.2 \times 10^{-5}\). The F-ratio is about 12, which may exceed the critical F at 95% (with 2 and 2 degrees of freedom, \(F_{\text{crit}} \approx 19\)). Here it passes, but this illustrates why the check matters.
Step 3 — Lack-of-fit test: The pure error variance from replicates is small. The lack-of-fit mean square is computed from the deviation of each level's mean from the regression line. The resulting F is, say, 1.8, well below the critical value of ~3.7 (for 4 and 12 degrees of freedom at 95%). Conclusion: No significant lack-of-fit — the linear model is adequate.
Step 4 — Report: Slope = 0.1503, intercept = 0.0015, \(R^2 = 0.9997\), lack-of-fit F = 1.8 (not significant). The calibration curve is linear over 0–5 mg/L.
---
Common Pitfalls
- Relying only on R². A curved dataset can still yield R² > 0.99. Always run the lack-of-fit test.
- Ignoring heteroscedasticity. If variance grows with concentration, a weighted regression or a transformation may be needed before linearity can be assessed.
- Removing outliers without cause. Outlier tests flag statistical anomalies; you must have a technical reason (e.g., injection error) to exclude a point.
- Extrapolating beyond the calibrated range. Linearity is only claimed within the tested concentration interval.
- Using too few levels. Three points can always be fit by a straight line; ISO 8466-1 requires at least five or six levels for a meaningful lack-of-fit test.
---
Closing
Calibration curve linearity is not a single number — it is a statistical verdict based on variance checks, outlier tests, and lack-of-fit analysis, as prescribed by ISO 8466-1 and ICH Q2. To apply this method quickly and correctly, try the free Calibration Curve Linearity calculator at https://www.6sq.com/tools/calcurve/ — it guides you through the ISO 8466-1 steps and gives you the statistical verdict in seconds.
No related results found
Invited:
0 replies