Related Experiment Videos
Selection of useful predictors in multivariate calibration.
M Forina1, S Lanteri, M C Cerrato Oliveros
1Department of Pharmaceutical and Food Chemistry and Technology, University of Genova, Via Brigata Salerno (s/n), 16147, Genova, Italy. forina@dictfa.unige.it
Analytical and Bioanalytical Chemistry
|September 7, 2004
Summary
This study evaluates ten predictor selection techniques for multivariate regression, finding that improved stepwise ordinary least-squares (SOLS) and iterative predictors weighting (IPW) are efficient parsimonious methods, while Westad-Martens uncertainty test (MUT) and uninformative variables elimination (UVE) excel as conservative approaches.
Area of Science:
- Chemometrics
- Statistical Modeling
Background:
- Accurate predictor selection is crucial for reliable multivariate calibration and regression.
- Existing techniques vary in their ability to identify useful predictors and prevent overfitting.
Purpose of the Study:
- To evaluate and compare ten predictor selection techniques for multivariate regression.
- To assess techniques based on performance metrics like predictive power and ability to detect useless predictors.
Main Methods:
- Comparison of classical stepwise ordinary least-squares (SOLS), genetic algorithms, and partial least-squares (PLS) based methods.
- Evaluation using real and artificial data, categorizing techniques as conservative or parsimonious.
- Introduction of improved SOLS using F-statistic plots and random data comparison.
Main Results:
- Westad-Martens uncertainty test (MUT) and uninformative variables elimination (UVE) identified as efficient conservative techniques.
- Improved SOLS demonstrated as an effective parsimonious technique, reducing overfitting.
- Iterative predictors weighting (IPW) presented as an alternative for minimum predictor set selection.
Conclusions:
- The choice between conservative and parsimonious techniques depends on the goal: retaining all informative predictors or selecting a minimal sufficient set.
- Improved SOLS and IPW offer efficient parsimonious selection, while MUT and UVE are strong conservative options.
- External validation or complete validation is recommended to prevent overestimation of prediction ability.