A comparison of calibration methods based on calibration data size and robustness

Least squares (LS) regression, ridge regression (RR) and partial least squares (PLS) regression have been widely used in statistical calibration of near infrared (NIR) instruments. Comparison of these methods has attracted lots of interest in literature. However, most papers compare calibration methods on the basis of a single experiment and focus on 'accuracy' rather than 'robustness'. In 'real life', the average accuracy level of various calibration methods may not make that much of a difference, but having an extremely bad prediction may be unacceptable. As well, a calibration set may be very expensive to acquire and methods that work well on small calibration sets may be preferred. In this paper, we compare least squares regression, ridge regression and partial least squares regression in the context of the varying calibration data size. Three data sets used in this study are all NIR-based measurements of fat or protein in milk. For a given calibration data size, 100 simulation experiments are carried out, the average value and 95 percentile of the root mean squared prediction errors of each method are compared. We found that relative performance of the calibration methods depends on data set and calibration data size as well.