Robust regression used for the treatment of partial non-linearity in multivariate calibration

Robust regression is proposed for attacking the problem of partial non-linearity in multivariate calibration in this paper. The non-linear spectral wavelengths were first regarded as outliers deviated from the assumed linear model. In order to reduce or eliminate the influence of the nonlinear spectral wavelengths upon the estimation of the concentration different types of weighting factors used in robust regression were tested. The commonly used Huber-type, Hampel-type and Andrews-type M-estimators were adopted in the robust regression and their performances were compared. The results for numeric simulation and real analytical systems have shown that robust regression may cope with partial non-linearity favourably.