n this paper, a number of proper applications both of class-modelling and discriminant classification to identification issues with beer samples were presented: the modelling approach was applied for characterising Trappist and Rochefort beers while the discriminant approach allowed the differentiation of Rochefort 8° and Rochefort 10° beers; all of these operations were based on NIR transflectance spectral data.
In the characterisation of Trappist beers, class-models managed to achieve only moderate levels of sensitivity and specificity with SIMCA performing best. One reason for this mediocre performance may be, as discussed earlier, the very complex distribution of data in the multidimensional space being modelled; this is claimed to especially affect model specificity. UNEQ in particular is reported not to perform well when the variable distribution is not normal (Forina et al., 2008a). Another reason seems likely to be the inclusion in the complete dataset of variables which contain little or no information about the modelling problem. Further work with this dataset on the elimination of non-informative variables is therefore warranted. In relation to the Rochefort versus non-Rochefort problem, UNEQ produced the best models using confidence levels of 90%. Discriminant analysis for differentiating Rochefort 8° and Rochefort 10° beers provided valuable results. This may be due to the lower complexity of the reduced data set, containing only Rochefort beers, and to the fact that NIR spectra contain information well correlated to the alcohol content of samples, a significant difference between them.