COMPARISON OF VARIABLE SELECTION AND REGRESSION METHODS IN MULTIVARIATE CALIBRATION OF A PROCESS ANALYZER

Ordinary least squares (OLS), partial least squares (PLS) and principa l component regression (PCR) were compared in the calibration of a pro cess analyzer with and without variable selection. The leave-one-out c ross-validation procedure was used in selecting variables for the OLS, PLS and PCR models and for the PCR models with component selection. I n the comparison the five data sets measured from different processes were used. The data were highly collinear. Although the OLS method was clearly improved through variable selection the PLS regression and pr incipal component regression were the best methods in the calibration of four data sets. In the calibration of one data set, every method me ntioned above succeeded well.