Enhancing canonical variate analysis by taking the scaling effect into account

Abstract

Sensory profiling aims to describe the sensory characteristics of food products using a list of descriptors with a panel of trained assessors. During sensory evaluations (or any other scoring task), individual differences in the scale width effectively used by assessors (scaling effect) are regularly observed. This scaling effect was included in a statistical model, the Mixed Assessor Model (MAM). This scaling effect can be decomposed into a physiological (descriptor-specific scaling) component and a psychological (overall scaling) component.The present paper shows how to take into account both physiological and psychological scaling effects in the Canonical Variate Analysis (CVA) framework. Agreement ellipses representing the pure disagreement of the subjects (scaling effect removed) are plotted around the product means. Thus, the differences between two products can be assessed by comparing their distances to the size of their agreement ellipses.Our so-called “overall CVA” and “MAM−CVA” method were compared to CVA and Principal Component Analysis (PCA) on 334 datasets. The sensory interpretations were similar for all maps but more differences between products were observed with MAM−CVA and overall CVA.An R package is offered to produce the maps presented in this paper.