ASSESSMENT OF THE QUALITY OF LATENT VARIABLE CALIBRATIONS BASED ON MONTE-CARLO SIMULATIONS

A general problem in multivariate calibration is the assessment of the prediction quality in advance, i.e., without extensive experimentatio n. Monte Carlo simulations are proposed to estimate the quality of lat ent variable calibrations and thereby to minimize experimental work an d to save resources. Two different data sets from industrial practice illustrate that this approach can be applied successfully for very dif ferent problems. The first set consists of 100 samples of penicillin a nalyzed by near-IR spectroscopy. The second data set contains visible spectra of 37 samples of a dyestuff intermediate. Based on the knowled ge of the noise in the calibration method, the noise in the reference method, the spectra of the analytes, the concentration range used for calibration, and the number of calibration samples, the prediction qua lity can be estimated. To characterize the quality of a multivariate c alibration, the simultaneous use of criteria related to the standard e rror of prediction (SEP) and correlation based criteria is recommended .