Investigating relationships in sensory and instrumental data using component-based methods
This talk focuses on new ways for investigating sensory-instrumental data relationships using component-based methods. The proposed approaches are related to manuscripts that are in preparation at the time of this abstract submission. The proposed methods could be used in a wide range of studies, but the talk will emphasize studies involving one reference treatment and many test treatments, which is relatively common. For instance, the reference treatment could be an in-market product, a gold-standard product, or a product at time 0 in a shelf-life study; test treatments could be prototypes, formulations, or treatments measured at different time points in a shelf-life study. Typically, many instrumental variables (e.g. chemical concentrations) are measured, but not all instrumental variables are associated with the attributes measured (e.g. aroma and flavor intensities) in sensory evaluation. The goal is to explain and predict sensory attribute intensities from the instrumental data. In the case of test-control paired comparisons, analytical measures associated with (and perhaps responsible for) differences in how the control and test treatments are perceived are usually of primary interest. How should these data relationships be investigated? Multiple linear regression yields unstable regression estimates due to multicollinearity in explanatory variables, which is why the regression tends to be solved using component-based methods. This talk proposes new approaches for investigating data relationships using supervised principal component regression (SPCR) and partial least squares regression (PLSR) based on a subset of paired comparison results. Methods are described briefly. Some key findings are stated. SPCR of all objects is equivalent to SPCR of all paired comparisons, but different from SPCR of test-control paired comparisons. PLSR of all objects is equivalent to PLSR of all paired comparisons, but different from PLSR of test-control paired comparisons. Models based on all objects and all paired comparisons explain the data relationships equivalently but differ in prediction accuracy measures. Implications are discussed. The talk will conclude with methodological recommendations. In studies where test-control paired comparisons are of primary interest, analyses should focus on these test-control paired comparisons, which yield different regression coefficients. This original presentation (not presented elsewhere; contains ideas that will be new to nearly everyone) is called a “review” because its purpose is to draw attention to the proposed methods by providing an entertaining high-level summary of the key ideas.
Castura, J.C. (2024). Investigating relationships in sensory and instrumental data using component-based methods. Society of Sensory Professionals 2024 Conference. 2-4 October. Pittsburgh, PA, USA. (Oral).