Selection and weighting of samples in multivariate regression model updating

Multivariate regression models are valid only for prediction of samples that are within the range of calibration data. Prediction of dependent variables for samples carrying new sources of variance requires updating of the model. A simple and convenient way to extend the model is to re-calibrate it using new incoming samples. However, it may be difficult and expensive to collect a high number of new samples and to analyse their property of interest with the reference method. Consequently, it is important to know what is the minimal number of samples necessary to update efficiently the model. A possibility would be to use only very few samples and to give them more weight, e.g. by including several copies of them. In this work, the impact of weighting on the performance of updated models is studied. For data sets studied, the weight applied to samples used for the update of the model has less importance than the number of these samples and their representativity, i.e. representative methods of selection of samples lead to better results than the other ones.