Robust and non parametric statistics in the validation of chemical analysis methods

Different robust statistical methodologies are applied to the evaluation of the repeatability and concentration-different intermediate precision expec ted in several procedures (polarographic and spectroscopic determinations w ith univariate and multivariate analysis of data) used in the determination of benzaldehyde at different concentrations between 0.02 and 0.19 mM. The robust estimation of the dispersion of the relative errors calculated using the Huber-estimator H15 for different concentration levels avoids the prac tical problem of detecting outlier data and prevents the underestimations t hat their elimination causes. Further, a regression of the calculated concentration versus the true one i s made for each procedure incorporating in the analysis the outlier data de tection by means of the Least Median of Squares Regression. LMS. This strat egy provides an estimation, without outlier data, of a common dispersion co rresponding to the range of concentrations studied. In this paper, this dis persion is compared to the one obtained directly by the robust estimation o f the repeatability. In addition this regression allows one to check the tr ueness of each procedure. According to the results obtained with the different approximations propose d for precision, it can be concluded that the multivariate procedures (Part ial Least Squares regression, PLS) are more precise, being of minor dispers ion the one which uses spectrophotometric measurements.