BUCHI NIRCal - Manual (Page 55)

Chemometrics
NIRCal 5.5 Manual, Version A 55
If two parameters are not systematically correlated, the mathematical approximation of the spectra via
PLS can never be performed for both parameters when each parameter is calibrated using its own
PLS. For this reason, it is recommended to calculate properties which are not systematically correlated
(e.g. ethanol and acetone contents in any given solvent mixture) in separate calibrations.
Therefore, whereas PCR reduces spectral data to the most dominate dimensions, the PLS aims at the
most relevant dimensions (relevant here means: best match between predicted and original values).
With PLS, the PC's are calculated exactly in relation to the highest correlation in the first PC.
3.2 Calibration Validation Methods
3.2.1 Validation Set (VS)
C-Set and V-Set
C-Set (Calibration Set).
From all spectra within a project only the spectra which are in the C-Set are used for the calculation of
the calibration.
V-Set (Validation Set).
From all spectra within a project only the spectra which are in the V-Set are used for an internal
validation of the calibration.
Normally for VS mode 2/3 of the spectra are selected as C-Set and 1/3 are in the V-Set. This can be
done with Toolbox “Set Creation”.
3.2.2 Cross Validation (CV)
Cross Validation Method
Instead of dividing the samples into two groups, a calibration set and a validation set, all samples are
used for calibration in CV mode. Several calibration runs are performed with all samples except a
small group with which the actual calibration is tested. This group is changed for all trial runs. The
validation results of all the runs are stored and lead to a standard error of the cross validation (SECV)
which compares well with the standard error of predictions (SEP) obtained in VS-mode.
In case each CV group consists of only one sample (one leave out) the method is called full cross
validation. This is the default setting for CV-grouping in NIRCal.
Cross validation is recommended for calibrations based on a small amount of samples. If this number
is larger than 50, NIRCal suggests to use a validation set instead (this comment can be suppressed;
see NIRCal Configuration Edit / Options / Calibration Defaults).
NOTE
All spectra of one sample must be assigned to the same CV-group.
NOTE
For CV mode all spectra should be assigned to the C-Set.
Limitations:

Cross validation is available for PCR and PLS only

Cross validation requires at least 4 spectra in the C-Set

For cross validation at least 2 CV groups have to be assigned

Cross validation will delete the V-Set assignment