Answers are grounded in this paper and cite their sources.
Evaluates impact of critical value methods on Q and T2 statistics in PLS-CM, suggesting improved modeling accuracy.
Two important statistics (T 2 , squared Mahalanobis distance, and Q, Euclidean distance) must be considered when applying PLS because their critical values define the boundaries within which the model can be applied (PLSbox). The calculation of the critical values of these statistics is based on the assumption that they follow a parametric probability distribution, usually F or χ 2 . These distributions were originally proposed to test hypotheses about the mean vector of a multivariate normal distribution and were later adapted for Principal Component Analysis. Subsequently, they were transferred without any discussion to the new bilinear decomposition induced by PLS models in the X -block Partial Least Squares Class Modeling, PLS-CM, is a soft two-class compliant classification method that builds a model for each class so that a sample can be assigned to one, both, or none of them, with the PLSbox playing a crucial role in the process. Through the analysis of five real cases, this work proves the inadequacy of the three most commonly used parametric estimators of Q and T 2 distributions, along with their impact on the PLSbox and thus, on the assignation of samples performed by PLS-CM. The proposal involves a Gaussian kernel for fitting the Q and T 2 distributions, and therefore, the critical values that define the PLSbox. The assessment of PLS-CM is conducted through a figure of merit more sensitive than the total efficiency: Diagonally Modified Confusion Entropy. It is shown that, unlike the usual approaches, no lack of fit occurs for both statistics in the datasets studied.
No takes yet. Share an insight, caveat, or question.
Valencia et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: