BUCHI NIRCal - Manual (Page 128)

NIRCal 5.5 Software Manual
128 NIRCal 5.5 Manual, Version A
Select the number of PCs, where:

the in the plot regression coefficients[1] the PCs, which have similar constant value, are good
for the calibration, big deviation indicates over fitting;

the SEP Generalized Cross Validation is small (about the value of the standard deviation of
lab method);

the V-Set Bias is around zero;

the Q-Value is high;

the absolute value of the PCR B-Matrix is high (not available by PLS);

the V- and C-Set regression coefficients are as close to one as possible;

the consistency is around 100;

the V- and C-Set PRESS are as small as possible;

the V- and C-Set SEP and SEE (SEC) are as small as possible and are similar (consistency).
Summarising all the information available from the graphics for this example, the first 3 principal
components (1-3) should be used as secondary PCs.
It is possible that different numbers of PCs are ideal for different selection criteria. In this situation the
different secondary PCs should be adjusted, the calibrations recalculated and the results compared.
In general for the C-Set a higher number of PCs always improve the result. For the V-Set, after a
certain number of PCs the result can be even worse. The optimum should be selected.
The selected number of secondary PCs should be adjusted and the calibration recalculated.
Note: The PLS algorithm calcu
lates the PC‘s with the highest correlation to the property values, that
means the secondary PC selection can not have a gap. The real selection is always "1 to the last
selected". For 1-2, 4-
7 secondary PC‘s selection the internal used secondary PC`s are 1-7 due to the
PLS algorithm. This allows easy switching between the methods form PCR to PLS and back without
losing the PC selection.