Cluster trimming: Improving segment homogeneity in product testing with consumers

Sensory and consumer researchers are often required to identify clusters of consumers based on product liking. A widespread, if informally held, belief is that many studies contain a group of consumers whose data differ from the product means in their assigned cluster. The implication could be less discrimination between samples within a cluster, leading to poorer decision-making. This research presents a variant of existing cluster analysis methods to potentially improve insights by excluding subjects with non-typical response behaviour. The approach is intuitive, easy to implement, and cluster algorithm agnostic. After removing non-discriminators (i.e., those who like all samples equally), the principle is to conduct cluster analysis and then calculate the correlation of each subject's liking ratings with the mean in their assigned cluster, setting a fixed “trim out” threshold for the correlation coefficient. Consumers with correlations below the threshold are classified as non-typical and excluded. After introducing cluster trimming in a case study, the method is applied to another 29 data sets representing category appraisal and product optimisation applications. On average, the cluster-trimming process resulted in around 1/3 of study participants being “trimmed out.” These consumers tended to discriminate less between samples than the participants who were retained following cluster trimming. While cluster trimming was found to add some value regarding product discrimination, this benefit was somewhat obscured by the reduction in cluster size. Exploring within-cluster heterogeneity, made possible by examining histograms of correlation coefficients (r-values), is recommended. Because the cluster trimming approach is correlational, it should not be used in product studies with too few samples ( n < 6). • A correlational approach for refinement of consumer segments based on liking. • Subjects are retained or trimmed out based on similarity with the cluster product means. • Improves cluster homogeneity but has little impact on sample discrimination. • Cluster trimming is easy to perform and does not bias mean liking ratings.