Analysis of the beverage data using cluster analysis, rotated principal components analysis and LOESS curves

Abstract

In an attempt to determine what influences liking of the 28 beverages in the study, a combination of data analysis techniques was employed. While statistical rigor is an essential component of all sensory and consumer data analyses, these analyses would be remiss without the partnership of the expertise of both a sensory and statistical professional. This paper will discuss how specific statistical methodologies have been guided by a combination of solutions that provide the most acceptable statistical approach and the most interpretable sensory solution. Cluster analysis was used to determine the existence of consumer subgroups or clusters based on the liking patterns across the product set. Two consumer studies were conducted and for each, two distinct subgroups were identified. Cross tabulation of consumer subgroups with that of respondent demographics was employed to identify differences relevant to the subgroups. Rotated principal component analysis was used to create the sensory dimensions that describe/differentiate the product set. As a result six sensory dimensions were identified. Using the principal component scores, the products were grouped into nine flavor profile groups. LOESS curve fitting in a multi-panel display was used to relate the mean product liking scores of each consumer subgroup to the sensory component scores. This format allows for: (1) the determination of the nature of each sensory dimension's effect on liking for the subgroups in each of the two studies, and (2) an understanding of why the liking pattern for the subgroups differ. Through this it will be shown that the sensory dimensions effect on product liking differs for each of the subgroups in both of the studies. Finally, consumer attribute ratings were correlated with the sensory component scores to gain an understanding of the consumers' description of the product set.