A method for validation of reference sets in SIMCA modelling

A method for validation of the reference set in Soft Independent Modelling of Class Analogies (SIMCA) is proposed. The reference set is used to build the SIMCA model and the remaining samples are fitted to this model. Thus, it is important that the reference set is representative for the reference class. In this work it is suggested that the reference set can be validated by the jackknife procedure. The jackknife estimate of standard error for the reference set is determined by successively leaving one sample out. It is proposed that the standard error should be minimised for an optimal reference set. Minimisation of the standard error should be balanced with the loss of variation span for the reference set to avoid a too narrow reference class. The reference sets are optimised by changing the composition of the reference set. The suggested validation method is tested on two data sets from environmental monitoring surveys of oil fields in the North Sea.