Article,
Robust Calibration
Comprehensive Chemometrics, (2009)
DOI: DOI: 10.1016/B978-044452701-1.00080-6
Abstract
Most multivariate statistical methods which are typically used for calibration, such as least-squares regression and PLSR, all rely on sample means and covariance matrices. These estimators are however very sensitive to possible outliers in the data, and consequently they can provide very bad fts and diagnostics about the data. Robust statistical methods try to find a fit which is similar to the fit one would obtain with-out the outliers. Moreover they allow to identify the outliers by their large residuals from that robust fit. In this chapter robust procedures are described for multivariate location and covariance estimation, and linear calibration in low dimensions. Next, the corresponding calibration methods for high-dimensional data (PCA, PCR, PLSR) are presented, as well as robustness in multi-way data. Also robust classification, and robustness of support vector machines is discussed.
