Development of robust multivariate calibration models

Multivariate calibration is frequently used for the quantitative analysis of a wide variety of materials using spectroscopy in the agricultural and food industries, manufacturing industries, medical sciences, and pharmaceutical industries. To date, most of the research activities in the multivariate calibration literature have focused on data analysis for model building and prediction with methods such as principal-components regression or partial least squares regression. As an alternative to focusing on data-analytic activities, we consider the ability of an experimental design to improve the robustness of the resulting calibration model. Through an example involving diffuse reflectance measurements, we illustrate how consideration of environmental and instrumental factors during the experimental design phase can result in a calibration model that is robust (against natural environmental and instrument variations) and easy to maintain. In this example, the analyte of interest produces a spectroscopic signal that is very weak in comparison to the environmental and instrumental factors.