INCORPORATING AUXILIARY PREDICTOR VARIATION IN PRINCIPAL COMPONENT REGRESSION-MODELS

In analytical chemistry a single fitted calibration model is used repe atedly to predict the level of the analyte of interest for the specime ns comprising the prediction set. Unlike the calibration (or training) set, which is often limited in size, the prediction set can be very l arge. In the case of multivariate calibration a number of methods such as PLS and PCR are commonly used to construct the calibration model. The set of instrumental measurements and the reference analyte level a re available for each specimen in the calibration set. For specimens i n the prediction set, only the instrumental measurements are available , since the problem is to predict the analyte level for these specimen s. It is not widely recognized that predictions of the analyte levels for individual specimens can be improved by utilizing seemingly unrela ted information from the instrumental measurements associated with the other members of the prediction set. In the case of PCR there exists a very straightforward procedure for doing this. A description of the various sources of prediction errors is provided to explain the abilit y of PCR to utilize this additional information. The use of PCR in thi s context is illustrated with both a synthetic and a real example.