Multivariate Bayesian discrimination for varietal authentication of Chilean red wine

Abstract

The process through which food or beverages is verified as complying with its label description is called food authentication. We propose to treat the authentication process as a classification problem. We consider multivariate observations and propose a multivariate Bayesian classifier that extends results from the univariate linear mixed model to the multivariate case. The model allows for correlation between wine samples from the same valley. We apply the proposed model to concentration measurements of nine chemical compounds named anthocyanins in 399 samples of Chilean red wines of the varieties Merlot, Carménère and Cabernet Sauvignon, vintages 2001–2004. We find satisfactory results, with a misclassification error rate based on a leave-one-out cross-validation approach of about 4%. The multivariate extension can be generally applied to authentication of food and beverages, where it is common to have several dependent measurements per sample unit, and it would not be appropriate to treat these as independent univariate versions of a common model.