Frequently a multivariate regression model for instrument calibration is de
veloped oil a master instrument and also used on other (server) instruments
, ri typical example is in Near Infrared Spectroscopy (NIRS). Because of th
e difference between the instruments (nature of the instrument, spectrophot
ometer or filter, number of wavelengths for spectrophotometer, lamp age. en
vironment) the regression model developed on the master instrument can't be
applied (different number of prediction) with the predictors measured on t
hr slave instrument.
Many procedures have been suggested. Frequently. an attempt to correct the
results (predicted response) of the server instrument by means of a linear
correction equation (slope + bias correction, SEC) is made. at least before
trying to use more complex procedures. sometimes this correction is reduce
d to the simple bias correction.
SBC is based on the hypothesis of a linear relationship between the respons
e value as predicted by the use of the master model on the server instrumen
t) and the true value. In principle it can be applied independently on the
other characteristics of the slave instrument, as the number of predictors.
In this paper a procedure has been developed to evaluate the validity of sl
ope-bias correction by using only the N samples used for the regression mod
el development. whose spectra are recorded on both the instruments. The pre
diction of the response is performed in full-validation (FV). In each FV cy
cle a sample is in the external set. The regression model is developed with
its optimum PLS complexity evaluated in predictive leave-one-out optimisat
ion. The regression model is used on the sample in the external set using e
ither the master spectrum and the Fen er one. After the regression, the eva
luation of the correction procedure is made with reference to the before re
gression standard deviation and to the standard deviations of the models de
veloped on the two instruments.
The suggested procedure has the advantage of using a reduced number of samp
les analysed with the reference technique, it seems to work well, except wh
en the quality of the two instruments is too different, so that the transfe
r performance is limited by the worse instrument.
Results obtained with simulated data and with several real data sets are pr
esented.