Principal component analysis versus canonical variate analysis. For the analysis of sensory profiles. Meta-analysis of 387 sensory datasets

International audiencePrincipal Component Analysis (PCA) of product mean scores is generally used to generate a product map from sensory profiling data. This approach does not take into account variance of these product mean scores due to individual variability. Canonical Variate Analysis (CVA) of the product effect in the two-way (product*panelist) multivariate ANOVA model is the natural extension of the classical univariate approach. This analysis generates successive components maximizing the ANOVA F-criterion. However, CVA requests the inversion of a covariance matrix which can result in computing instability when the sensory attributes are highly correlated. The paper compares results from covariance PCA and CVA, both performed on significant (p=0.05) attributes, and assesses severity of attribute colinearities based on the analysis of 387 descriptive sensory studies gathered in SensoBase (www.sensobase.fr). Although colinearity diagnostics were significant one time out of two, product configurations from PCA and CVA were found very similar (average RV of 0.94). However, product discrimination based on first and second components, measured by the MANOVA-F ratio, was in average 1.9 times higher in CVA than PCA. In overall, these data suggest that having used PCA extensively in sensory analysis did not result in misinterpretation and loss of information. However, the gain of discrimination, brought by CVA, certainly results in product maps with smaller confidence ellipses, thus easier to interpret. The same investigation will be performed with no selection of significant attributes, by type of attributes (visual, texture, flavor …) and on correlation rather covariance PCA, in order to delineate whether these alternative ways of analyzing descriptive data modify former conclusions or not