Working the Analyte X Curve

A six-point calibration for a photometric method, Analyte X, reported in mg/dL, runs six calibrators with assigned values of 0, 50, 100, 150, 200, and 250 mg/dL and records an absorbance signal (unitless) for each: 0.021, 0.144, 0.273, 0.392, 0.518, and 0.647. The software fits a first-order linear regression to these six signal-response pairs.

The fitted line is signal = 0.0024977 × concentration + 0.020286. That equation is the working curve: from this point on, every patient specimen's absorbance is converted back to a concentration using this line, until the next calibration.

Back-calculation checks the fit by running each calibrator's own signal back through the equation and comparing the result to its assigned value. At the 250 mg/dL level, the observed signal of 0.647 back-calculates to 250.9 mg/dL, a residual of plus 0.9 mg/dL. At the 0 mg/dL level, the residual is plus 0.3 mg/dL. Across all six points, the residuals scatter between about minus 1.2 and plus 1.2 mg/dL, with no consistent upward or downward trend as concentration increases.

That pattern, small, random scatter around zero with no trend, is what supports accepting this curve. A residual pattern that grows steadily larger at one end, or curves rather than scatters, would instead signal nonlinearity or a proportional bias that a straight-line fit is absorbing into its slope rather than reporting as error, the situation the opening scenario's high-level failure resembles. CLSI EP06 frames this same logic for a full linearity study: judge acceptability from the deviation at each level against a predefined, clinically meaningful limit, not from a summary statistic like R-squared alone, because a high R-squared can coexist with a real, clinically relevant curve or bias.

A good calibration curve is not just a high R-squared, it is small, patternless residuals at every level you checked, including the ones near the edges of the range.