Investigation on reference value error in multivariate calibration

Abstract

Near infrared spectral analysis is a kind of indirect measuring technology, that is, calibration model is established by using sample spectra and the reference value of the interested concentration, and then predict the interested concentration of the unknown sample by using its spectrum and calibration model. Generally, the quality of calibration model is evaluated by the deviation between the predicted value and the reference value of the measured concentration. Reference value measurement error of the calibration set samples and model error are the main error source of multivariate calibration. When there was reference value measurement error, the regression model would cause great prediction deviation. However in practical application, apparent prediction error caused by reference value measurement error is always considered as the faultiness of calibration model. So it is falsely regarded that the prediction error of multivariate calibration model would be greater than that of reference method. In this paper, effect of reference value measurement error on prediction error was analyzed, and the relationship among apparent prediction error, actual prediction error and reference value measurement error was discussed. At the same time, the applicability and rationality of a computation formula was also discussed, which could estimate the true prediction error in a more precise way. It was proved that the prediction accuracy of an excellent calibration model was greater than that of the reference method by analyzing the reference value measurement error. An experiment of glucose aqueous solution samples was given in this paper. In this example, the effect of reference value measurement error on the prediction error of multivariate calibration model was investigated, and the estimating formula of the corrected prediction error was evaluated. The conclusion could be universally applied to not only multivariate calibration MLR, PCR, PLS and so on, but also in much more complex samples.