Investigating sensory-instrumental relationships in a subset of test-control paired comparisons by partial least squares regression
Relationships between sensory and instrumental variables are often investigated using partial least squares regression (PLSR). We describe the conventional PLSR solution. Next, we describe two new approaches for conducting PLSR based on paired comparisons of objects. First, we describe the PLSR of all paired comparisons. Conducting PLSR of all paired comparisons is equivalent to the conventional solution in the sense that both analyses yield the same regression coefficients. Next, we consider a case where multiple test formulations are compared with one control formulation. In this case, the test-control paired comparisons are of primary interest and the test-test paired comparisons are of lesser interest. Conducting PLSR of the relevant subset of paired comparisons yields regression coefficients that differ from the conventional PLSR solution. Compared to the conventional PLSR solution, the PLSR of the relevant subset of paired comparisons explains the test-control paired differences well. However, the PLSR of the relevant subset of paired comparisons does not always give better prediction accuracy. Findings are illustrated using a published data set including instrumental and sensory evaluation of French cabernet franc wines.
Castura, J.C., Tomic, O., & Næs, T. (2024). Investigating sensory-instrumental relationships in a subset of test-control paired comparisons by partial least squares regression. AgroStat 2024 Conference. 3-6 September. Bragança, Portugal. (Oral Presentation).