Recent advances in the estimation of multivariate and multiway analytical figures of merit

OLIVIERI, A. C.

San Pablo

Congreso; 46th World Chemistry Congress; 2017

With a growing number of analytical methodologies incorporating multivariate calibration models, the development of suitable estimators for their analytical figures of merit is playing an important role in modern analytical chemistry. The main goal of this presentation is to discuss the recent trends regarding reliable and interpretable estimators of traditional figures of merit such as sensitivity, prediction uncertainty and detection limit.In the case of first-order multivariate calibration, heteroscedasticity in the instrumental measurements is driving some of the recent developments, and influencing the definition of new estimators. Prediction uncertainty for generalized error structures (correlated and/or heteroscedastic) has been recently described.1 This led to a new proposal concerning the sensitivity parameter for first-order calibration, because the traditional one was shown to be inappropriate for describing the analytical performance in the presence of general noise structures.2 Future research efforts in the field will be definitely oriented to the unbiased and efficient estimation of error covariance matrices, which are needed to define correct figures of merit for multivariate systems.Advances in prediction uncertainty inevitable lead to new insights into the detection capabilities of multivariate calibration. Recently, a IUPAC-consistent approach has been taken to the definition of the limit of detection in partial least-squares calibration3 and in neural network calibration based on radial basis functions.4 These two developments may have important consequences in the field of near infrared spectroscopic analysis of intact materials, a subject of utmost importance in many industrial areas.Regarding multiway calibration, i.e., when the data show various physical dimensions or modes, a general scheme for estimating sensitivity, prediction uncertainty and detection limit has been published recently.5 In the presence of heteroscedastic and correlated instrumental noise, it is apparent that new definitions of analytical figures of merit are required, in line with the analogous developments in first-order multivariate counterparts.6This body of work complements the relevant development of multivariate and multiway calibration models that have taken place in the last decades, honoring the famous statement: ?an analytical result is not (yet) a result, unless accompanied by an uncertainty estimate?.