Effect of scaling regimes on the prediction of analytical results from multivariate calibration

Abstract

Multivariate calibration has become indispensable for many analytical methods which suffer from insufficient selectivity. One of the practical limitations of multivariate calibration is the unfavourable error propagation for minor and trace components. For partial least-squares regression (PLS), it is shown that the accuracy and ruggedness of results depend critically on the scaling of the concentration matrix. When the scales for each chemical component are changed separately, the weighting schemes obtained reduce the relative error for minor components without negative effects on the results for major components. The conclusions drawn from simulation studies were confirming by testing mixtures of organic dyes (chromazurol S, phenyl red, bromocresol green and 2-(2-pyridylazo)-1-naphthol-4-sulphonic acid).