For multivariate calibration models such as those used on process IR analyzers, sample (matrix) specific biases can be a significant portion of the model error. These calibration models are typically developed on large numbers of samples covering a wide range of composition and process variation. Initial validation is done on only 20 samples which span a far smaller range of composition and process variation. As a result, the inherent sample (matrix) specific biases can lead to an apparent bias during initial validation. While the bias may be significant relative to the variation of these initial results, it may well be within the expected range of sample specific biases seen during the calibration. The D6708 based validation needs to be reexamined to take into account sample (matrix) specific biases inherent to these types models to provide better recommendations on interpretation of results and actions to take based on results. Additionally, the current practices assume that the spectral and primary test method measurements are done on the same material. Currently, applications have developed where the spectral measurement is conducted on a blend stock, whereas the primary test method is done on a final blended product. The multivariate model predicts the property of the blended product based on the spectrum of the blend stock. The scope of the practice needs to be updated to include these newer applications.
Keywordscontrol chart; infrared analyzer; infrared spectrophotometers; IR spectroscopy; multivariate process; NIR spectroscopy; statistical quality assurance; validation;