Metrohm Vision – Tutorial - Manual (Page 30)

1.7 Partial Least Squares Regression 
NIRS Vision Tutorial – Operation 
23
 Correlation:
Generally, the selected wavelength should be where the absolute value of
correlation is high.
When using 2nd derivative spectra, locate an area of high negative
correlation. (When using raw spectra, use a wavelength with high positive
correlation.) The wavelength should be selected in a region where
correlation is relatively broad and flat, not narrow and spike-like.
 Sensitivity:
Selecting a wavelength with low sensitivity makes the calibration more
robust, because small changes in absorbance will not result in large relative
changes in the reported constituent value.
 Statistical values:
High F-value, low K(1), R 2 close to 1, small SEC around 1-1.5 times Laberror
of reference method.
Edit Equation
1. Click on Regression  Edit Wavelength
2. In the Edit Wavelength dialog box enter:
 Wavelength 1675
 Click on [OK]
1.7 Partial Least Squares Regression
Partial Least Squares (PLS) regression allows a user to use the spectral
information of a whole range of wavelengths, even the full spectrum. This
method eliminates colinearity (high correlation between wavelength terms)
that is difficult to overcome with classical methods.
1.7.1 Sample Selection
A second Product, HiPro, has been chosen for developing a Partial Least
Squares Regression. Sample Selection must be performed on this dataset prior
to developing the equation.