Analyzing Product and Individual Differences in Sensory Discrimination Testing by Thurstonian and Statistical models

Abstract

Sensory discrimination tests are used to gain information about products by using the human senses to evaluate the samples. More specifically, a sensory discrimination study is conducted when the the aim is to investigate whether products are perceptibly different. Such studies are often considered for food, beverages as well as personal care products. An example is when a company gets a new supplier of an ingredient in one of their products. It is of high importance to investigate how this change of the ingredient affects the product. Even though the chemical composition of the product changes, it does not necessarily mean that people can detect the difference. These days, people become more and more interested in how to improve their health. This is also reflected in the companies desire to make their products healthier without changing how the product is perceived by their customers. Therefore, it is important to conduct sensory discrimination tests when ingredients are changed. This thesis is concerned with the analysis of product and individual differences in sensory discrimination testing. Sensory discrimination tests become more and more advanced raising a need for new types of analysis of sensory discrimination data. This thesis contributes with the development of Thurstonian models and how these can be aligned with well-known statistical models. Generalized linear mixed models are used in many applications. However, it is not common to consider such complicated models when considering sensory discrimination tests. Actually, sensory discrimination tests are often analyzed by too simplistic methods, ignoring important variables, such as individuals, that affect the results of the analysis. One focus of this project is to propose a way to incorporate such effects in the models when analyzing data from sensory discrimination studies. These models, including random effects, are called Thurstonian mixed models. Considering generalized linear mixed models for sensory discrimination studies opens up for many possibilities. It becomes possible to gain information about the individuals, the so-called assessors, as well as making more proper conclusions regarding the products. Moreover, the estimates of product and individual differences are obtained on the d-prime scale. Often multiple sensory attributes are considered in a discrimination study. These can be analyzed individually by the Thurstonian mixed models we are introducing. This thesis is presenting a multivariate analysis to gain knowledge about the product and individual differences across the sensory attributes. This is achieved by analyzing the product and individual differences, on the d-prime scale, by principal component analysis. Sensory discrimination tests are sometimes conducted to investigate the performance of sensory panels or to compare different laboratories. In such tests, multiple d-prime values can be obtained. For sensory discrimination tests, which lead to binomially distributed responses, we propose a new test statistic for the comparison of multiple d-prime values. The test statistic we suggest is an improved way of analyzing multiple d-prime values compared to a previous suggested test statistic.