Which just-about-right feature should be changed if evaluations deviate? A case study using sum of ranking differences

Abstract

Several different approaches have been introduced for analysis of just-about-right (JAR) data; however, their results are sometimes deviating or even contradictory. More reliable results are gained, if a consensus of many methods is determined. A specific approach is presented to compare and select JAR attributes of food products. Overall liking was set as dependent (Y) variable and the JAR variables were used as independent (X) variables for regression methods. The mean drop value, difference between the mean overall liking of the attribute as optimum (or JAR) and that of mean overall liking of the attribute as an extreme, i.e. too much and not enough, was used for penalty analysis and its variants. Generalized Pair Correlation Method (GPCM) compares the impact of the JAR variables on overall liking pairwise and the probability weighted difference ordering was applied for ordering the attributes. A special data fusion is suggested based on the sum of ranking differences (SRD), primarily developed for method comparison. SRD method was able to rank the JAR variables based on their differences from a benchmark defined by all of the JAR evaluation methods in maximal performance. This enables also to group the product attributes. Moreover, it gives recommendations for how to optimize the products based on the results of several JAR methods and helps to gain a more reliable evaluation and selection of JAR attributes. The significant features can be identified easily when the SRD procedure is completed by the frequencies of consumer evaluations. The same data matrix transposed is suitable to rank the evaluation methods using the average of all evaluation methods (consensus). From among the JAR evaluation techniques, GPCM proved to be closest to the average, i.e. it can be used for substitution of the other techniques.