Konferenzbeitrag
Robust Linear Models and Soft Modelling to Explain Chemical and Environmental Data
Vorschaubild nicht verfügbar
Volltext URI
Dokumententyp
Text/Conference Paper
Zusatzinformation
Datum
2005
Autor:innen
Zeitschriftentitel
ISSN der Zeitschrift
Bandtitel
Verlag
Masaryk University Brno
Zusammenfassung
Both analytical and physical data measured in chemistry and environment have very often non-normal distribution, which makes the determination of true limits and risk by Gauss distribution curve impossible. Linear regression models are commonly used to construct prediction models to estimate unknown (e.g. future) properties based on known chemical or other physical properties. Often, the data used to construct such prediction models does not come from a pre designed experiment and may contain outliers or nontypical data, which can completely spoil the regression model when classical least squares method is used to estimate the regression parameters. We used a BIR (Bounded Influence Regression) to estimate parameters from uneven and noisy data with outliers. This approach resulted to better estimates with lower variances and to disclosure of outliers that was difficult to identify by classical methods. A solution of the big errors risk in the data are robust regression methods, that take into account possibility of such errors and treat them using weights, iterative re-calculations, or use of different criteria than least squares. For real data we recommend a simplified procedure to use both approaches – classical least squares and robust method. If the results are the same or similar, use least squares, if the results are different, trust rather robust method than least squares. Some graphical diagnostic methods are also mentioned. The procedures are available in a program QCE for Windows.